use pre-commit

.github/workflows/publish.yml

@@ -26,15 +26,15 @@ jobs:
         uses: actions/setup-python@v5
         with:
           python-version: "3.8"
-      
+
       - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
           python -m pip install build
-      
+
       - name: Build package
         run: |
           python -m build
-      
+
       - name: Publish package
         uses: pypa/gh-action-pypi-publish@release/v1

data/alpaca_en_demo.json

@@ -4999,4 +4999,4 @@
     "input": "Time waits for no one.",
     "output": "No one can stop time from moving forward."
   }
-]
+]

data/alpaca_zh_demo.json

@@ -4999,4 +4999,4 @@
     "input": "",
     "output": "安第斯山脉位于南美洲，横跨七个国家，包括委内瑞拉，哥伦比亚，厄瓜多尔，秘鲁，玻利维亚，智利和阿根廷。安第斯山脉是世界上最长的山脉之一，全长约7,000千米（4,350英里），其山脉沿着南美洲西海岸蜿蜒延伸，平均海拔约为4,000米（13,000英尺）。在其南部，安第斯山脉宽度达到700千米（430英里），在其北部宽度约为500千米（310英里）。"
   }
-]
+]

data/belle_multiturn/belle_multiturn.py

@@ -17,9 +17,9 @@ _CITATION = """\
 }
 """
 
-_HOMEPAGE = "{}/datasets/BelleGroup/multiturn_chat_0.8M".format(_HF_ENDPOINT)
+_HOMEPAGE = f"{_HF_ENDPOINT}/datasets/BelleGroup/multiturn_chat_0.8M"
 _LICENSE = "gpl-3.0"
-_URL = "{}/datasets/BelleGroup/multiturn_chat_0.8M/resolve/main/multiturn_chat_0.8M.json".format(_HF_ENDPOINT)
+_URL = f"{_HF_ENDPOINT}/datasets/BelleGroup/multiturn_chat_0.8M/resolve/main/multiturn_chat_0.8M.json"
 
 
 class BelleMultiturn(datasets.GeneratorBasedBuilder):
@@ -38,7 +38,7 @@ class BelleMultiturn(datasets.GeneratorBasedBuilder):
         return [datasets.SplitGenerator(name=datasets.Split.TRAIN, gen_kwargs={"filepath": file_path})]
 
     def _generate_examples(self, filepath: str):
-        with open(filepath, "r", encoding="utf-8") as f:
+        with open(filepath, encoding="utf-8") as f:
             for key, row in enumerate(f):
                 data = json.loads(row)
                 conversations = []

data/c4_demo.json

@@ -899,4 +899,4 @@
   {
     "text": "Having a DisplayPort with you at your house will always keep you entertained if you love gaming and live to stream. The cables fit perfectly in a monitor, projector, laptop or tablet. There are many products in the market you can choose from, but you have to do proper research to avoid purchasing a wrong product.\nThis is the perfect cable for your laptop or desktop that is equipped with a display port. You can also use it to enjoy your movies where you can also expand your laptop for better view or when you are presenting at school. It provides you with quality and high-definition audio and videos from your laptop.\nThis cable is gold plated, and this helps to resist corrosion or resist any damage and enhance durability. This kind of plating also ensures a perfect signal transmission. This display port connector provides a secure connection where it has a release button that is always lowered before unplugging. The display port is also portable where you can carry it anywhere with you.\nThis mini display port cable can give you quality transmission of videos and audios. You can connect it to your computer or tablet to get HD videos. Has gold plated connectors, copper conductors, and foil-and-braid shielding where all these provide enough energy and convenient connection.\nThis cable measures a maximum of 10 feet and gives an easy time when connecting it to your laptop. The cable has a sleek design that gives you an easy time to carry and store the cable. It is secure and reliable where it will provide you with a good connection when you need it.\nIf you are looking for a perfect cable for video streaming and gaming, this is the perfect fit. It has a construction design that prevents the cable from cracking or any damage. It contains gold-plated connectors and multiple shielding where this prevents any damage and enhance the durability of the cable.\nThe cable provides you with free non-interfered data transmission from any time and position. Provides a secure and reliable connection where you only plug and play without struggling. You can directly connect the display port that is equipped to your desktop for high definition videos and clear audios.\nThis is a convenient cable that connects a display port cable directly to a projector or HD monitor with display port unit. This cable is portable and mobile where you can use it anywhere. It features gold-plated connectors that with copper connectors and braid shielding where they all combine to provide powerful cable performance reliable connections.\nIt transmits high definition videos and clear music from your computer to a monitor for gaming or video streaming. It provides a secure connection where it has a button that is always lowered before unplugging. This cable is 10 foot long, and this gives you an easy time when using it from any position.\nThis cable has gold-plated connectors that resist corrosion and any form of damage. The gold connectors also enhance reliable and increased connectivity. It features inner braided foil that reduces interference and increases video and audio quality. The 6 feet cable connects a display port enabled computer to an HD monitor or projector with DP port unit for audio and video streaming.\nThe cable gives you the best and quality video resolution you need. The DP has a latch that provides a secure and reliable connection with the port where the latch must be pressed down before unplugging the connector. This is perfect for extended display or mirrored displays.\nThe cable supports resolutions of up to 3840 by 2160 and high-definition videos. This display port provides you with a secure and reliable connection any time you need it. It has pure copper conductors and triple metal shielding that enhance a perfect performance.\nThis cable transmits both high definition video from your computer or tablet to an HD display. It has gold-plated and copper conductors that protect the cable from damage and improve its performance.\nHas foil and braid shielding that helps in reducing electromagnetic interference.\nIt has an ergonomic design where it has secured grips for easy plugging and unplugging. The cable measures 6 foot and supports a high video resolution. This cable provides you with a high-quality connection no matter where you are.\nThis is a perfect choice when you need quality and clear videos and pictures. The cable provides you with a high video resolution. The cable is 10 foot long to provide you with enough and secure connection from any position.\nThe cable is gold-plated for preventing corrosion and providing a secure connection. Has a release button that is always lowered before unplugging. Has a slim design where you can fold it easily for easy carrying.\nThis cable has latching gold-plated connectors for providing a secure and reliable connection. The gold plating also helps in resisting damage, and this enhances the durability of the cable. Features braided copper conductors for providing a reliLine truncated
   }
-]
+]

data/dataset_info.json

@@ -625,4 +625,4 @@
     },
     "folder": "python"
   }
-}
+}

data/dpo_en_demo.json

@@ -7223,4 +7223,4 @@
       "value": "Abstraction is a principle in object-oriented programming that refers to the process of focusing on the essential features of an object or concept without emphasizing the details that are not important for its functionality. It enables programmers to create classes that represent the objects they want to work with, and specifies only the behavior of those objects, leaving out unnecessary details. Abstraction helps to make the code more maintainable, modular, and scalable. It also improves the productivity of developers by reducing the amount of code they need to write."
     }
   }
-]
+]

data/dpo_zh_demo.json

@@ -5055,4 +5055,4 @@
       "value": "C. 参与讨论"
     }
   }
-]
+]

data/glaive_toolcall_en_demo.json

@@ -9155,4 +9155,4 @@
     ],
     "tools": "[]"
   }
-]
+]

data/glaive_toolcall_zh_demo.json

@@ -9019,4 +9019,4 @@
     ],
     "tools": "[]"
   }
-]
+]

data/hh_rlhf_en/hh_rlhf_en.py

@@ -8,9 +8,9 @@ import datasets
 _HF_ENDPOINT = os.getenv("HF_ENDPOINT", "https://huggingface.co")
 _DESCRIPTION = "Human preference data about helpfulness and harmlessness."
 _CITATION = ""
-_HOMEPAGE = "{}/datasets/Anthropic/hh-rlhf".format(_HF_ENDPOINT)
+_HOMEPAGE = f"{_HF_ENDPOINT}/datasets/Anthropic/hh-rlhf"
 _LICENSE = "mit"
-_URL = "{}/datasets/Anthropic/hh-rlhf/resolve/main/".format(_HF_ENDPOINT)
+_URL = f"{_HF_ENDPOINT}/datasets/Anthropic/hh-rlhf/resolve/main/"
 _URLS = {
     "train": [
         _URL + "harmless-base/train.jsonl.gz",
@@ -53,7 +53,7 @@ class HhRlhfEn(datasets.GeneratorBasedBuilder):
     def _generate_examples(self, filepaths: List[str]):
         key = 0
         for filepath in filepaths:
-            with open(filepath, "r", encoding="utf-8") as f:
+            with open(filepath, encoding="utf-8") as f:
                 for row in f:
                     data = json.loads(row)
                     chosen = data["chosen"]

data/identity.json

@@ -454,4 +454,4 @@
     "input": "",
     "output": "抱歉，我不是 OpenAI 开发的 ChatGPT，我是 {{author}} 开发的 {{name}}，旨在为用户提供智能化的回答和帮助。"
   }
-]
+]

data/kto_en_demo.json

@@ -5395,4 +5395,4 @@
     ],
     "label": false
   }
-]
+]

data/mllm_demo.json

@@ -137,4 +137,4 @@
       "mllm_demo_data/3.jpg"
     ]
   }
-]
+]

data/mllm_video_demo.json

@@ -44,4 +44,4 @@
       "mllm_demo_data/3.mp4"
     ]
   }
-]
+]

data/ultra_chat/ultra_chat.py

@@ -20,9 +20,9 @@ _CITATION = """\
 }
 """
 
-_HOMEPAGE = "{}/datasets/stingning/ultrachat".format(_HF_ENDPOINT)
+_HOMEPAGE = f"{_HF_ENDPOINT}/datasets/stingning/ultrachat"
 _LICENSE = "cc-by-nc-4.0"
-_BASE_DATA_URL = "{}/datasets/stingning/ultrachat/resolve/main/train_{{idx}}.jsonl".format(_HF_ENDPOINT)
+_BASE_DATA_URL = f"{_HF_ENDPOINT}/datasets/stingning/ultrachat/resolve/main/train_{{idx}}.jsonl"
 
 
 class UltraChat(datasets.GeneratorBasedBuilder):
@@ -42,7 +42,7 @@ class UltraChat(datasets.GeneratorBasedBuilder):
 
     def _generate_examples(self, filepaths: List[str]):
         for filepath in filepaths:
-            with open(filepath, "r", encoding="utf-8") as f:
+            with open(filepath, encoding="utf-8") as f:
                 for row in f:
                     try:
                         data = json.loads(row)

data/wiki_demo.txt

@@ -27,4 +27,4 @@ ASCII ( ), abbreviated from American Standard Code for Information Interchange,
 Austin is the capital of Texas in the United States.Austin may also refer to:Geographical locationsAustralia Austin, Western AustraliaCanada Austin, Manitoba Austin, Ontario Austin, Quebec Austin Island, NunavutFrance Saint-Austin, hamlet at la Neuville-Chant-d'Oisel, NormandyHong Kong  Austin station (MTR), KowloonUnited States Austin, Arkansas Austin, Colorado Austin Township, Macon County, Illinois Austin, Chicago, Cook County, Illinois Austin, Indiana Austin, Kentucky Austin, Minnesota Austin, Missouri Austin, Nevada Austin, Ohio Austin, Oregon Austin, Pennsylvania Austin, Texas Austin County, Texas (note that the city of Austin, Texas is located in Travis County)Schools Austin College, Sherman, Texas University of Texas at Austin, flagship institution of the University of Texas System Austin Peay State University, Clarksville, TennesseeReligion Augustine of Hippo An adjective for the AugustiniansBusiness American Austin Car Company, short-lived American automobile maker Austin Automobile Company, short-lived American automobile company Austin Motor Company, British car manufacturer Austin cookies and crackers, Keebler Company brandEntertainment "Austin" (song), a single by Blake Shelton Austin, a kangaroo Beanie Baby produced by Ty, Inc. Austin the kangaroo from the children's television series The BackyardigansOther uses Austin (building), a building designed by artist Ellsworth Kelly under construction in Austin, Texas Austin (given name), a short form of Augustin, or Augustine, including fictional characters Austin (surname) USS Austin, three shipsSee also All pages beginning with Austin August (disambiguation) Augustin (disambiguation) Augustine (disambiguation) Austin station (disambiguation) Austins (disambiguation) Austen (disambiguation) Justice Austin (disambiguation) Austinburg (disambiguation)
 Animation is a method in which figures are manipulated to appear as moving images. In traditional animation, images are drawn or painted by hand on transparent celluloid sheets to be photographed and exhibited on film. Today, most animations are made with computer-generated imagery (CGI). Computer animation can be very detailed 3D animation, while 2D computer animation (which may have the look of traditional animation) can be used for stylistic reasons, low bandwidth, or faster real-time renderings. Other common animation methods apply a stop motion technique to two- and three-dimensional objects like paper cutouts, puppets, or clay figures.An animated cartoon is an animated film, usually a short film aimed at children and featuring an exaggerated visual style. The style takes inspiration from comic strips, often featuring anthropomorphic animals, superheroes, or the adventures of child protagonists. Especially with animals that form a natural predator/prey relationship (e.g. cats and mice, coyotes and birds) the action often centers around violent pratfalls such as falls, collisions and explosions that would be lethal in real life.  Commonly, animators achieved the effect by a rapid succession of images that minimally differ from each other. The illusion—as in motion pictures in general—is thought to rely on the phi phenomenon and beta movement, but the exact causes are still uncertain. Analog mechanical animation media that rely on the rapid display of sequential images include the phénakisticope, zoetrope, flip book, praxinoscope, and film. Television and video are popular electronic animation media that originally were analog and now operate digitally. For display on computers, technology such as the animated GIF and Flash animation were developed.In addition to short films, feature films, television series, animated GIFs, and other media dedicated to the display of moving images, animation is also prevalent in video games, motion graphics, user interfaces, and visual effects.The physical movement of image parts through simple mechanics—for instance moving images in magic lantern shows—can also be considered animation. The mechanical manipulation of three-dimensional puppets and objects to emulate living beings has a very long history in automata. Electronic automata were popularized by Disney as animatronics.EtymologyThe word "animation" stems from the Latin "animātiōn", stem of "animātiō", meaning "a bestowing of life". The primary meaning of the English word is "liveliness" and has been in use much longer than the meaning of "moving image medium".HistoryBefore cinematographyHundreds of years before the introduction of true animation, people all over the world enjoyed shows with moving figures that were created and manipulated manually in puppetry, automata, shadow play, and the magic lantern. The multi-media phantasmagoria shows that were very popular in European theatres from the late 18th century through the first half of the 19th century, featured lifelike projections of moving ghosts and other frightful imagery in motion.In 1833, the stroboscopic disc (better known as the phénakisticope) introduced the principle of modern animation with sequential images that were shown one by one in quick succession to form an optical illusion of motion pictures. Series of sequential images had occasionally been made over thousands of years, but the stroboscopic disc provided the first method to represent such images in fluent motion and for the first time had artists creating series with a proper systematic breakdown of movements. The stroboscopic animation principle was also applied in the zoetrope (1866), the flip book (1868) and the praxinoscope (1877). A typical 19th-century animation contained about 12 images that were displayed as a continuous loop by spinning a device manually. The flip book often contained more pictures and had a beginning and end, but its animation would not last longer than a few seconds. The first to create much longer sequences seems to have been Charles-Émile Reynaud, who between 1892 and 1900 had much success with his 10- to 15-minute-long Pantomimes Lumineuses.Silent eraWhen cinematography eventually broke through in 1895 after animated pictures had been known for decades, the wonder of the realistic details in the new medium was seen as its biggest accomplishment. Animation on film was not commercialized until a few years later by manufacturers of optical toys, with chromolithography film loops (often traced from live-action footage) for adapted toy magic lanterns intended for kids to use at home. It would take some more years before animation reached movie theaters.After earlier experiments by movie pioneers J. Stuart Blackton, Arthur Melbourne-Cooper, Segundo de Chomón, and Edwin S. Porter (among others), Blackton's The Haunted Hotel (1907) was the first huge stop motion success, baffling audiences by showing objects that apparently moved by themselves in fullLine truncated
 Apollo is one of the  Olympian deities in classical Greek and Roman religion and Greek and Roman mythology. The national divinity of the Greeks, Apollo has been recognized as a god of archery, music and dance, truth and prophecy, healing and diseases, the Sun and light, poetry, and more. One of the most important and complex of the Greek gods, he is the son of Zeus and Leto, and the twin brother of Artemis, goddess of the hunt. Seen as the most beautiful god and the ideal of the kouros (ephebe, or a beardless, athletic youth), Apollo is considered to be the most Greek of all the gods. Apollo is known in Greek-influenced Etruscan mythology as Apulu.As the patron deity of Delphi (Apollo Pythios), Apollo is an oracular god—the prophetic deity of the Delphic Oracle. Apollo is the god who affords help and wards off evil; various epithets call him the "averter of evil". Delphic Apollo is the patron of seafarers, foreigners and the protector of fugitives and refugees.Medicine and healing are associated with Apollo, whether through the god himself or mediated through his son Asclepius. Apollo delivered people from epidemics, yet he is also a god who could bring ill-health and deadly plague with his arrows. The invention of archery itself is credited to Apollo and his sister Artemis. Apollo is usually described as carrying a golden bow and a quiver of silver arrows. Apollo's capacity to make youths grow is one of the best attested facets of his panhellenic cult persona. As the protector of young (kourotrophos), Apollo is concerned with the health and education of children. He presided over their passage into adulthood. Long hair, which was the prerogative of boys, was cut at the coming of age (ephebeia) and dedicated to Apollo.Apollo is an important pastoral deity, and was the patron of herdsmen and shepherds. Protection of herds, flocks and crops from diseases, pests and predators were his primary duties. On the other hand, Apollo also encouraged founding new towns and establishment of civil constitution. He is associated with dominion over colonists. He was the giver of laws, and his oracles were consulted before setting laws in a city.As the god of mousike, Apollo presides over all music, songs, dance and poetry. He is the inventor of string-music, and the frequent companion of the Muses, functioning as their chorus leader in celebrations. The lyre is a common attribute of Apollo. In Hellenistic times, especially during the 5th century BCE, as Apollo Helios he became identified among Greeks with Helios, the personification of the sun. In Latin texts, however, there was no conflation of Apollo with Sol among the classical Latin poets until 1st century CE. Apollo and Helios/Sol remained separate beings in literary and mythological texts until the 5th century CE.EtymologyApollo (Attic, Ionic, and Homeric Greek: , Apollōn ( ); Doric: , Apellōn; Arcadocypriot: , Apeilōn; Aeolic: , Aploun; )The name Apollo—unlike the related older name Paean—is generally not found in the Linear B (Mycenean Greek) texts, although there is a possible attestation in the lacunose form ]pe-rjo-[ (Linear B: ]-[) on the KN E 842 tablet, though it has also been suggested that the name might actually read "Hyperion" ([u]-pe-rjo-[ne]).The etymology of the name is uncertain. The spelling  ( in Classical Attic) had almost superseded all other forms by the beginning of the common era, but the Doric form, Apellon (), is more archaic, as it is derived from an earlier . It probably is a cognate to the Doric month Apellaios (), and the offerings apellaia () at the initiation of the young men during the family-festival apellai (). According to some scholars, the words are derived from the Doric word apella (), which originally meant "wall," "fence for animals" and later "assembly within the limits of the square." Apella () is the name of the popular assembly in Sparta, corresponding to the ecclesia (). R. S. P. Beekes rejected the connection of the theonym with the noun apellai and suggested a Pre-Greek proto-form *Apalyun.Several instances of popular etymology are attested from ancient authors. Thus, the Greeks most often associated Apollo's name with the Greek verb  (apollymi), "to destroy". Plato in Cratylus connects the name with  (apolysis), "redemption", with  (apolousis), "purification", and with  ([h]aploun), "simple", in particular in reference to the Thessalian form of the name, , and finally with  (aeiballon), "ever-shooting". Hesychius connects the name Apollo with the Doric  (apella), which means "assembly", so that Apollo would be the god of political life, and he also gives the explanation  (sekos), "fold", in which case Apollo would be the god of flocks and herds. In the ancient Macedonian language  (pella) means "stone," and some toponyms may be derived from this word:  (Pella, the capital of ancient Macedonia) and  (Pellēnē/Pellene).A number of non-Greek etymologies have been suggested for the name, The Hittite form ApaliunLine truncated
-Andre Kirk Agassi ( ; born April 29, 1970) is an American former world No. 1 tennis player. He is an eight-time major champion and a 1996 Olympic gold medalist, as well as a runner-up in seven other Grand Slam tournaments.Agassi was the first man to win four Australian Open singles titles in the Open Era (though later surpassed by Novak Djokovic, who won his fifth title in 2015 and has since won the tournament nine times). Agassi is the second of five men to achieve the career Grand Slam in the Open Era and the fifth of eight overall to make the achievement. He is also the first of two men to achieve the career Golden Slam (career Grand Slam and Olympic gold medal), and the only man to win a career Super Slam (career Grand Slam, plus the Olympic gold medal and the year-end championships).Agassi was the first man to win all four singles majors on three different surfaces (hard, clay and grass), and remains the most recent American man to win the French Open (in 1999) and the Australian Open (in 2003). He also won 17 ATP Masters Series titles and was part of the winning Davis Cup teams in 1990, 1992 and 1995. Agassi reached the world No. 1 ranking for the first time in 1995 but was troubled by personal issues during the mid-to-late 1990s and sank to No. 141 in 1997, prompting many to believe that his career was over. Agassi returned to No. 1 in 1999 and enjoyed the most successful run of his career over the next four years. During his 20-plus year tour career, Agassi was known by the nickname "The Punisher".After suffering from sciatica caused by two bulging discs in his back, a spondylolisthesis (vertebral displacement) and a bone spur that interfered with the nerve, Agassi retired from professional tennis on September 3, 2006, after losing in the third round of the US Open. He is the founder of the Andre Agassi Charitable Foundation, which has raised over $60 million for at-risk children in Southern Nevada. In 2001, the Foundation opened the Andre Agassi College Preparatory Academy in Las Vegas, a K–12 public charter school for at-risk children. He has been married to fellow tennis player Steffi Graf since 2001.1970–1985: Early lifeAndre Agassi was born in Las Vegas, Nevada, to Emmanuel "Mike" Agassi, a former Olympic boxer from Iran and American Elizabeth "Betty" Agassi (née Dudley). His father is of Armenian and Assyrian heritage. Andre Agassi's mother, Betty, is a breast cancer survivor. He has three older siblings – Rita (last wife of former number one Pancho Gonzales), Philip and Tami. Andre was given the middle name Kirk after Kirk Kerkorian, an Armenian American billionaire. Emmanuel Agassi, then a waiter at Tropicana Las Vegas, had met Kerkorian in 1963.At the age of 12, Agassi and his good friend and doubles partner, Roddy Parks, won the 1982 National Indoor Boys 14s Doubles Championship in Chicago. Agassi describes memorable experiences and juvenile pranks with Roddy in his book Open.When he was 13, Agassi was sent to Nick Bollettieri's Tennis Academy in Florida. He was meant to stay for only three months, because that was all his father could afford. After thirty minutes of watching Agassi play, Bollettieri, deeply impressed by his talent, called Mike and said: "Take your check back. He's here for free." Agassi then dropped out of school in the ninth grade to pursue a full-time tennis career.1986–2006: Professional career1986–1993: Breakthrough and the first major titleAgassi turned professional at the age of 16 and competed in his first tournament at La Quinta, California. He won his first match against John Austin, but then lost his second match to Mats Wilander. By the end of 1986, Agassi was ranked No. 91. He won his first top-level singles title in 1987 at the Sul American Open in Itaparica and ended the year ranked No. 25. He won six additional tournaments in 1988 (Memphis, U.S. Men's Clay Court Championships, Forest Hills WCT, Stuttgart Outdoor, Volvo International and Livingston Open), and, by December of that year, he had surpassed US$1 million in career prize money after playing in just 43 tournaments—the fastest anyone in history had reached that level. During 1988, he also set the open-era record for most consecutive victories by a male teenager (a record that stood for 17 years until Rafael Nadal broke it in 2005). His year-end ranking was No. 3, behind second-ranked Ivan Lendl and top-ranked Mats Wilander. Both the Association of Tennis Professionals and Tennis magazine named Agassi the Most Improved Player of the Year for 1988.In addition to not playing the Australian Open (which later became his best Grand Slam event) for the first eight years of his career, Agassi chose not to play at Wimbledon from 1988 through 1990 and publicly stated that he did not wish to play there because of the event's traditionalism, particularly its "predominantly white" dress code to which players at the event are required to conform.Strong performances on the tour meant that Agassi wasLine truncated
+Andre Kirk Agassi ( ; born April 29, 1970) is an American former world No. 1 tennis player. He is an eight-time major champion and a 1996 Olympic gold medalist, as well as a runner-up in seven other Grand Slam tournaments.Agassi was the first man to win four Australian Open singles titles in the Open Era (though later surpassed by Novak Djokovic, who won his fifth title in 2015 and has since won the tournament nine times). Agassi is the second of five men to achieve the career Grand Slam in the Open Era and the fifth of eight overall to make the achievement. He is also the first of two men to achieve the career Golden Slam (career Grand Slam and Olympic gold medal), and the only man to win a career Super Slam (career Grand Slam, plus the Olympic gold medal and the year-end championships).Agassi was the first man to win all four singles majors on three different surfaces (hard, clay and grass), and remains the most recent American man to win the French Open (in 1999) and the Australian Open (in 2003). He also won 17 ATP Masters Series titles and was part of the winning Davis Cup teams in 1990, 1992 and 1995. Agassi reached the world No. 1 ranking for the first time in 1995 but was troubled by personal issues during the mid-to-late 1990s and sank to No. 141 in 1997, prompting many to believe that his career was over. Agassi returned to No. 1 in 1999 and enjoyed the most successful run of his career over the next four years. During his 20-plus year tour career, Agassi was known by the nickname "The Punisher".After suffering from sciatica caused by two bulging discs in his back, a spondylolisthesis (vertebral displacement) and a bone spur that interfered with the nerve, Agassi retired from professional tennis on September 3, 2006, after losing in the third round of the US Open. He is the founder of the Andre Agassi Charitable Foundation, which has raised over $60 million for at-risk children in Southern Nevada. In 2001, the Foundation opened the Andre Agassi College Preparatory Academy in Las Vegas, a K–12 public charter school for at-risk children. He has been married to fellow tennis player Steffi Graf since 2001.1970–1985: Early lifeAndre Agassi was born in Las Vegas, Nevada, to Emmanuel "Mike" Agassi, a former Olympic boxer from Iran and American Elizabeth "Betty" Agassi (née Dudley). His father is of Armenian and Assyrian heritage. Andre Agassi's mother, Betty, is a breast cancer survivor. He has three older siblings – Rita (last wife of former number one Pancho Gonzales), Philip and Tami. Andre was given the middle name Kirk after Kirk Kerkorian, an Armenian American billionaire. Emmanuel Agassi, then a waiter at Tropicana Las Vegas, had met Kerkorian in 1963.At the age of 12, Agassi and his good friend and doubles partner, Roddy Parks, won the 1982 National Indoor Boys 14s Doubles Championship in Chicago. Agassi describes memorable experiences and juvenile pranks with Roddy in his book Open.When he was 13, Agassi was sent to Nick Bollettieri's Tennis Academy in Florida. He was meant to stay for only three months, because that was all his father could afford. After thirty minutes of watching Agassi play, Bollettieri, deeply impressed by his talent, called Mike and said: "Take your check back. He's here for free." Agassi then dropped out of school in the ninth grade to pursue a full-time tennis career.1986–2006: Professional career1986–1993: Breakthrough and the first major titleAgassi turned professional at the age of 16 and competed in his first tournament at La Quinta, California. He won his first match against John Austin, but then lost his second match to Mats Wilander. By the end of 1986, Agassi was ranked No. 91. He won his first top-level singles title in 1987 at the Sul American Open in Itaparica and ended the year ranked No. 25. He won six additional tournaments in 1988 (Memphis, U.S. Men's Clay Court Championships, Forest Hills WCT, Stuttgart Outdoor, Volvo International and Livingston Open), and, by December of that year, he had surpassed US$1 million in career prize money after playing in just 43 tournaments—the fastest anyone in history had reached that level. During 1988, he also set the open-era record for most consecutive victories by a male teenager (a record that stood for 17 years until Rafael Nadal broke it in 2005). His year-end ranking was No. 3, behind second-ranked Ivan Lendl and top-ranked Mats Wilander. Both the Association of Tennis Professionals and Tennis magazine named Agassi the Most Improved Player of the Year for 1988.In addition to not playing the Australian Open (which later became his best Grand Slam event) for the first eight years of his career, Agassi chose not to play at Wimbledon from 1988 through 1990 and publicly stated that he did not wish to play there because of the event's traditionalism, particularly its "predominantly white" dress code to which players at the event are required to conform.Strong performances on the tour meant that Agassi wasLine truncated

evaluation/ceval/mapping.json

@@ -207,4 +207,4 @@
     "name": "兽医学",
     "category": "STEM"
   }
-}
+}

evaluation/cmmlu/mapping.json

@@ -267,4 +267,4 @@
     "name": "世界宗教",
     "category": "Humanities"
   }
-}
+}

evaluation/mmlu/mapping.json

@@ -227,4 +227,4 @@
     "name": "world religions",
     "category": "Humanities"
   }
-}
+}

evaluation/mmlu/mmlu.py

@@ -158,5 +158,4 @@ class MMLU(datasets.GeneratorBasedBuilder):
         df = pd.read_csv(filepath, header=None)
         df.columns = ["question", "A", "B", "C", "D", "answer"]
 
-        for i, instance in enumerate(df.to_dict(orient="records")):
-            yield i, instance
+        yield from enumerate(df.to_dict(orient="records"))

examples/deepspeed/ds_z0_config.json

@@ -25,4 +25,4 @@
     "contiguous_gradients": true,
     "round_robin_gradients": true
   }
-}
+}

examples/deepspeed/ds_z2_config.json

@@ -25,4 +25,4 @@
     "contiguous_gradients": true,
     "round_robin_gradients": true
   }
-}
+}

examples/deepspeed/ds_z2_offload_config.json

@@ -29,4 +29,4 @@
     "contiguous_gradients": true,
     "round_robin_gradients": true
   }
-}
+}

examples/deepspeed/ds_z3_config.json

@@ -27,4 +27,4 @@
     "stage3_max_reuse_distance": 1e9,
     "stage3_gather_16bit_weights_on_model_save": true
   }
-}
+}

examples/deepspeed/ds_z3_offload_config.json

@@ -35,4 +35,4 @@
     "stage3_max_reuse_distance": 1e9,
     "stage3_gather_16bit_weights_on_model_save": true
   }
-}
+}

scripts/cal_flops.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 Microsoft Corporation and the LlamaFactory team.
 #
 # This code is inspired by the Microsoft's DeepSpeed library.

scripts/cal_lr.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 imoneoi and the LlamaFactory team.
 #
 # This code is inspired by the imoneoi's OpenChat library.
@@ -74,7 +73,7 @@ def calculate_lr(
     elif stage == "sft":
         data_collator = DataCollatorForSeq2Seq(tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX)
     else:
-        raise NotImplementedError("Stage does not supported: {}.".format(stage))
+        raise NotImplementedError(f"Stage does not supported: {stage}.")
 
     dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     valid_tokens, total_tokens = 0, 0

scripts/cal_mfu.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -100,7 +99,7 @@ def compute_device_flops(world_size: int) -> float:
     elif "4090" in device_name:
         return 98 * 1e12 * world_size
     else:
-        raise NotImplementedError("Device not supported: {}.".format(device_name))
+        raise NotImplementedError(f"Device not supported: {device_name}.")
 
 
 def calculate_mfu(
@@ -140,10 +139,10 @@ def calculate_mfu(
         "bf16": True,
     }
     if deepspeed_stage in [2, 3]:
-        args["deepspeed"] = "examples/deepspeed/ds_z{}_config.json".format(deepspeed_stage)
+        args["deepspeed"] = f"examples/deepspeed/ds_z{deepspeed_stage}_config.json"
 
     run_exp(args)
-    with open(os.path.join("saves", "test_mfu", "all_results.json"), "r", encoding="utf-8") as f:
+    with open(os.path.join("saves", "test_mfu", "all_results.json"), encoding="utf-8") as f:
         result = json.load(f)
 
     if dist.is_initialized():
@@ -157,7 +156,7 @@ def calculate_mfu(
         * compute_model_flops(model_name_or_path, total_batch_size, seq_length)
         / compute_device_flops(world_size)
     )
-    print("MFU: {:.2f}%".format(mfu_value * 100))
+    print(f"MFU: {mfu_value * 100:.2f}%")
 
 
 if __name__ == "__main__":

scripts/cal_ppl.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -100,7 +99,7 @@ def calculate_ppl(
             tokenizer=tokenizer, label_pad_token_id=IGNORE_INDEX, train_on_prompt=train_on_prompt
         )
     else:
-        raise NotImplementedError("Stage does not supported: {}.".format(stage))
+        raise NotImplementedError(f"Stage does not supported: {stage}.")
 
     dataloader = DataLoader(trainset, batch_size, shuffle=False, collate_fn=data_collator, pin_memory=True)
     criterion = torch.nn.CrossEntropyLoss(reduction="none")
@@ -125,8 +124,8 @@ def calculate_ppl(
     with open(save_name, "w", encoding="utf-8") as f:
         json.dump(perplexities, f, indent=2)
 
-    print("Average perplexity is {:.2f}".format(total_ppl / len(perplexities)))
-    print("Perplexities have been saved at {}.".format(save_name))
+    print(f"Average perplexity is {total_ppl / len(perplexities):.2f}")
+    print(f"Perplexities have been saved at {save_name}.")
 
 
 if __name__ == "__main__":

scripts/length_cdf.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -61,7 +60,7 @@ def length_cdf(
     for length, count in length_tuples:
         count_accu += count
         prob_accu += count / total_num * 100
-        print("{:d} ({:.2f}%) samples have length < {}.".format(count_accu, prob_accu, length + interval))
+        print(f"{count_accu:d} ({prob_accu:.2f}%) samples have length < {length + interval}.")
 
 
 if __name__ == "__main__":

scripts/llama_pro.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 Tencent Inc. and the LlamaFactory team.
 #
 # This code is inspired by the Tencent's LLaMA-Pro library.
@@ -40,7 +39,7 @@ if TYPE_CHECKING:
 
 
 def change_name(name: str, old_index: int, new_index: int) -> str:
-    return name.replace(".{:d}.".format(old_index), ".{:d}.".format(new_index))
+    return name.replace(f".{old_index:d}.", f".{new_index:d}.")
 
 
 def block_expansion(
@@ -76,27 +75,27 @@ def block_expansion(
     state_dict = model.state_dict()
 
     if num_layers % num_expand != 0:
-        raise ValueError("`num_layers` {} should be divisible by `num_expand` {}.".format(num_layers, num_expand))
+        raise ValueError(f"`num_layers` {num_layers} should be divisible by `num_expand` {num_expand}.")
 
     split = num_layers // num_expand
     layer_cnt = 0
     output_state_dict = OrderedDict()
     for i in range(num_layers):
         for key, value in state_dict.items():
-            if ".{:d}.".format(i) in key:
+            if f".{i:d}." in key:
                 output_state_dict[change_name(key, i, layer_cnt)] = value
 
-        print("Add layer {} copied from layer {}".format(layer_cnt, i))
+        print(f"Add layer {layer_cnt} copied from layer {i}")
         layer_cnt += 1
         if (i + 1) % split == 0:
             for key, value in state_dict.items():
-                if ".{:d}.".format(i) in key:
+                if f".{i:d}." in key:
                     if "down_proj" in key or "o_proj" in key:
                         output_state_dict[change_name(key, i, layer_cnt)] = torch.zeros_like(value)
                     else:
                         output_state_dict[change_name(key, i, layer_cnt)] = torch.clone(value)
 
-            print("Add layer {} expanded from layer {}".format(layer_cnt, i))
+            print(f"Add layer {layer_cnt} expanded from layer {i}")
             layer_cnt += 1
 
     for key, value in state_dict.items():
@@ -113,17 +112,17 @@ def block_expansion(
             torch.save(shard, os.path.join(output_dir, shard_file))
 
     if index is None:
-        print("Model weights saved in {}".format(os.path.join(output_dir, weights_name)))
+        print(f"Model weights saved in {os.path.join(output_dir, weights_name)}")
     else:
         index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
         with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
             json.dump(index, f, indent=2, sort_keys=True)
-        print("Model weights saved in {}".format(output_dir))
+        print(f"Model weights saved in {output_dir}")
 
     print("- Fine-tune this model with:")
-    print("model_name_or_path: {}".format(output_dir))
+    print(f"model_name_or_path: {output_dir}")
     print("finetuning_type: freeze")
-    print("freeze_trainable_layers: {}".format(num_expand))
+    print(f"freeze_trainable_layers: {num_expand}")
     print("use_llama_pro: true")
 
 

scripts/llamafy_baichuan2.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -63,16 +62,16 @@ def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetenso
             torch.save(shard, os.path.join(output_dir, shard_file))
 
     if index is None:
-        print("Model weights saved in {}".format(os.path.join(output_dir, WEIGHTS_NAME)))
+        print(f"Model weights saved in {os.path.join(output_dir, WEIGHTS_NAME)}")
     else:
         index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
         with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
             json.dump(index, f, indent=2, sort_keys=True)
-        print("Model weights saved in {}".format(output_dir))
+        print(f"Model weights saved in {output_dir}")
 
 
 def save_config(input_dir: str, output_dir: str):
-    with open(os.path.join(input_dir, CONFIG_NAME), "r", encoding="utf-8") as f:
+    with open(os.path.join(input_dir, CONFIG_NAME), encoding="utf-8") as f:
         llama2_config_dict: Dict[str, Any] = json.load(f)
 
     llama2_config_dict["architectures"] = ["LlamaForCausalLM"]
@@ -82,7 +81,7 @@ def save_config(input_dir: str, output_dir: str):
 
     with open(os.path.join(output_dir, CONFIG_NAME), "w", encoding="utf-8") as f:
         json.dump(llama2_config_dict, f, indent=2)
-    print("Model config saved in {}".format(os.path.join(output_dir, CONFIG_NAME)))
+    print(f"Model config saved in {os.path.join(output_dir, CONFIG_NAME)}")
 
 
 def llamafy_baichuan2(

scripts/llamafy_qwen.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -86,7 +85,7 @@ def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetenso
             elif "lm_head" in key:
                 llama2_state_dict[key] = value
             else:
-                raise KeyError("Unable to process key {}".format(key))
+                raise KeyError(f"Unable to process key {key}")
 
     weights_name = SAFE_WEIGHTS_NAME if save_safetensors else WEIGHTS_NAME
     shards, index = shard_checkpoint(llama2_state_dict, max_shard_size=shard_size, weights_name=weights_name)
@@ -98,18 +97,18 @@ def save_weight(input_dir: str, output_dir: str, shard_size: str, save_safetenso
             torch.save(shard, os.path.join(output_dir, shard_file))
 
     if index is None:
-        print("Model weights saved in {}".format(os.path.join(output_dir, weights_name)))
+        print(f"Model weights saved in {os.path.join(output_dir, weights_name)}")
     else:
         index_name = SAFE_WEIGHTS_INDEX_NAME if save_safetensors else WEIGHTS_INDEX_NAME
         with open(os.path.join(output_dir, index_name), "w", encoding="utf-8") as f:
             json.dump(index, f, indent=2, sort_keys=True)
-        print("Model weights saved in {}".format(output_dir))
+        print(f"Model weights saved in {output_dir}")
 
     return str(torch_dtype).replace("torch.", "")
 
 
 def save_config(input_dir: str, output_dir: str, torch_dtype: str):
-    with open(os.path.join(input_dir, CONFIG_NAME), "r", encoding="utf-8") as f:
+    with open(os.path.join(input_dir, CONFIG_NAME), encoding="utf-8") as f:
         qwen_config_dict: Dict[str, Any] = json.load(f)
 
     llama2_config_dict: Dict[str, Any] = OrderedDict()
@@ -135,7 +134,7 @@ def save_config(input_dir: str, output_dir: str, torch_dtype: str):
 
     with open(os.path.join(output_dir, CONFIG_NAME), "w", encoding="utf-8") as f:
         json.dump(llama2_config_dict, f, indent=2)
-    print("Model config saved in {}".format(os.path.join(output_dir, CONFIG_NAME)))
+    print(f"Model config saved in {os.path.join(output_dir, CONFIG_NAME)}")
 
 
 def llamafy_qwen(

scripts/loftq_init.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
 #
 # This code is based on the HuggingFace's PEFT library.
@@ -70,19 +69,19 @@ def quantize_loftq(
     setattr(peft_model.peft_config["default"], "base_model_name_or_path", os.path.abspath(output_dir))
     setattr(peft_model.peft_config["default"], "init_lora_weights", True)  # don't apply loftq again
     peft_model.save_pretrained(loftq_dir, safe_serialization=save_safetensors)
-    print("Adapter weights saved in {}".format(loftq_dir))
+    print(f"Adapter weights saved in {loftq_dir}")
 
     # Save base model
     base_model: "PreTrainedModel" = peft_model.unload()
     base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
     tokenizer.save_pretrained(output_dir)
-    print("Model weights saved in {}".format(output_dir))
+    print(f"Model weights saved in {output_dir}")
 
     print("- Fine-tune this model with:")
-    print("model_name_or_path: {}".format(output_dir))
-    print("adapter_name_or_path: {}".format(loftq_dir))
+    print(f"model_name_or_path: {output_dir}")
+    print(f"adapter_name_or_path: {loftq_dir}")
     print("finetuning_type: lora")
-    print("quantization_bit: {}".format(loftq_bits))
+    print(f"quantization_bit: {loftq_bits}")
 
 
 if __name__ == "__main__":

scripts/pissa_init.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 HuggingFace Inc. and the LlamaFactory team.
 #
 # This code is based on the HuggingFace's PEFT library.
@@ -54,7 +53,7 @@ def quantize_pissa(
         lora_alpha=lora_alpha if lora_alpha is not None else lora_rank * 2,
         lora_dropout=lora_dropout,
         target_modules=lora_target,
-        init_lora_weights="pissa" if pissa_iter == -1 else "pissa_niter_{}".format(pissa_iter),
+        init_lora_weights="pissa" if pissa_iter == -1 else f"pissa_niter_{pissa_iter}",
     )
 
     # Init PiSSA model
@@ -65,17 +64,17 @@ def quantize_pissa(
     setattr(peft_model.peft_config["default"], "base_model_name_or_path", os.path.abspath(output_dir))
     setattr(peft_model.peft_config["default"], "init_lora_weights", True)  # don't apply pissa again
     peft_model.save_pretrained(pissa_dir, safe_serialization=save_safetensors)
-    print("Adapter weights saved in {}".format(pissa_dir))
+    print(f"Adapter weights saved in {pissa_dir}")
 
     # Save base model
     base_model: "PreTrainedModel" = peft_model.unload()
     base_model.save_pretrained(output_dir, safe_serialization=save_safetensors)
     tokenizer.save_pretrained(output_dir)
-    print("Model weights saved in {}".format(output_dir))
+    print(f"Model weights saved in {output_dir}")
 
     print("- Fine-tune this model with:")
-    print("model_name_or_path: {}".format(output_dir))
-    print("adapter_name_or_path: {}".format(pissa_dir))
+    print(f"model_name_or_path: {output_dir}")
+    print(f"adapter_name_or_path: {pissa_dir}")
     print("finetuning_type: lora")
     print("pissa_init: false")
     print("pissa_convert: true")

scripts/test_toolcall.py

@@ -1,4 +1,3 @@
-# coding=utf-8
 # Copyright 2024 the LlamaFactory team.
 #
 # Licensed under the Apache License, Version 2.0 (the "License");

setup.py

@@ -20,7 +20,7 @@ from setuptools import find_packages, setup
 
 
 def get_version() -> str:
-    with open(os.path.join("src", "llamafactory", "extras", "env.py"), "r", encoding="utf-8") as f:
+    with open(os.path.join("src", "llamafactory", "extras", "env.py"), encoding="utf-8") as f:
         file_content = f.read()
         pattern = r"{}\W*=\W*\"([^\"]+)\"".format("VERSION")
         (version,) = re.findall(pattern, file_content)
@@ -28,7 +28,7 @@ def get_version() -> str:
 
 
 def get_requires() -> List[str]:
-    with open("requirements.txt", "r", encoding="utf-8") as f:
+    with open("requirements.txt", encoding="utf-8") as f:
         file_content = f.read()
         lines = [line.strip() for line in file_content.strip().split("\n") if not line.startswith("#")]
         return lines
@@ -61,7 +61,7 @@ extra_require = {
     "qwen": ["transformers_stream_generator"],
     "modelscope": ["modelscope"],
     "openmind": ["openmind"],
-    "dev": ["ruff", "pytest"],
+    "dev": ["pre-commit", "ruff", "pytest"],
 }
 
 
@@ -72,7 +72,7 @@ def main():
         author="hiyouga",
         author_email="hiyouga" "@" "buaa.edu.cn",
         description="Easy-to-use LLM fine-tuning framework",
-        long_description=open("README.md", "r", encoding="utf-8").read(),
+        long_description=open("README.md", encoding="utf-8").read(),
         long_description_content_type="text/markdown",
         keywords=["LLaMA", "BLOOM", "Falcon", "LLM", "ChatGPT", "transformer", "pytorch", "deep learning"],
         license="Apache 2.0 License",

src/api.py

@@ -25,7 +25,7 @@ def main():
     app = create_app(chat_model)
     api_host = os.environ.get("API_HOST", "0.0.0.0")
     api_port = int(os.environ.get("API_PORT", "8000"))
-    print("Visit http://localhost:{}/docs for API document.".format(api_port))
+    print(f"Visit http://localhost:{api_port}/docs for API document.")
     uvicorn.run(app, host=api_host, port=api_port)
 
 

src/llamafactory/api/app.py

@@ -130,5 +130,5 @@ def run_api() -> None:
     app = create_app(chat_model)
     api_host = os.environ.get("API_HOST", "0.0.0.0")
     api_port = int(os.environ.get("API_PORT", "8000"))
-    print("Visit http://localhost:{}/docs for API document.".format(api_port))
+    print(f"Visit http://localhost:{api_port}/docs for API document.")
     uvicorn.run(app, host=api_host, port=api_port)

src/llamafactory/api/chat.py

@@ -70,7 +70,7 @@ ROLE_MAPPING = {
 def _process_request(
     request: "ChatCompletionRequest",
 ) -> Tuple[List[Dict[str, str]], Optional[str], Optional[str], Optional["ImageInput"]]:
-    logger.info("==== request ====\n{}".format(json.dumps(dictify(request), indent=2, ensure_ascii=False)))
+    logger.info(f"==== request ====\n{json.dumps(dictify(request), indent=2, ensure_ascii=False)}")
 
     if len(request.messages) == 0:
         raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid length")
@@ -142,7 +142,7 @@ def _create_stream_chat_completion_chunk(
 async def create_chat_completion_response(
     request: "ChatCompletionRequest", chat_model: "ChatModel"
 ) -> "ChatCompletionResponse":
-    completion_id = "chatcmpl-{}".format(uuid.uuid4().hex)
+    completion_id = f"chatcmpl-{uuid.uuid4().hex}"
     input_messages, system, tools, image = _process_request(request)
     responses = await chat_model.achat(
         input_messages,
@@ -169,7 +169,7 @@ async def create_chat_completion_response(
             tool_calls = []
             for tool in result:
                 function = Function(name=tool[0], arguments=tool[1])
-                tool_calls.append(FunctionCall(id="call_{}".format(uuid.uuid4().hex), function=function))
+                tool_calls.append(FunctionCall(id=f"call_{uuid.uuid4().hex}", function=function))
 
             response_message = ChatCompletionMessage(role=Role.ASSISTANT, tool_calls=tool_calls)
             finish_reason = Finish.TOOL
@@ -193,7 +193,7 @@ async def create_chat_completion_response(
 async def create_stream_chat_completion_response(
     request: "ChatCompletionRequest", chat_model: "ChatModel"
 ) -> AsyncGenerator[str, None]:
-    completion_id = "chatcmpl-{}".format(uuid.uuid4().hex)
+    completion_id = f"chatcmpl-{uuid.uuid4().hex}"
     input_messages, system, tools, image = _process_request(request)
     if tools:
         raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Cannot stream function calls.")
@@ -229,7 +229,7 @@ async def create_stream_chat_completion_response(
 async def create_score_evaluation_response(
     request: "ScoreEvaluationRequest", chat_model: "ChatModel"
 ) -> "ScoreEvaluationResponse":
-    score_id = "scoreval-{}".format(uuid.uuid4().hex)
+    score_id = f"scoreval-{uuid.uuid4().hex}"
     if len(request.messages) == 0:
         raise HTTPException(status_code=status.HTTP_400_BAD_REQUEST, detail="Invalid request")
 

src/llamafactory/chat/chat_model.py

@@ -53,7 +53,7 @@ class ChatModel:
         elif model_args.infer_backend == "vllm":
             self.engine: "BaseEngine" = VllmEngine(model_args, data_args, finetuning_args, generating_args)
         else:
-            raise NotImplementedError("Unknown backend: {}".format(model_args.infer_backend))
+            raise NotImplementedError(f"Unknown backend: {model_args.infer_backend}")
 
         self._loop = asyncio.new_event_loop()
         self._thread = Thread(target=_start_background_loop, args=(self._loop,), daemon=True)

src/llamafactory/chat/vllm_engine.py

@@ -105,7 +105,7 @@ class VllmEngine(BaseEngine):
         video: Optional["VideoInput"] = None,
         **input_kwargs,
     ) -> AsyncIterator["RequestOutput"]:
-        request_id = "chatcmpl-{}".format(uuid.uuid4().hex)
+        request_id = f"chatcmpl-{uuid.uuid4().hex}"
         if image is not None:
             if IMAGE_PLACEHOLDER not in messages[0]["content"]:
                 messages[0]["content"] = IMAGE_PLACEHOLDER + messages[0]["content"]
@@ -159,7 +159,7 @@ class VllmEngine(BaseEngine):
 
         if image is not None:  # add image features
             if not isinstance(image, (str, ImageObject)):
-                raise ValueError("Expected image input is a path or PIL.Image, but got {}.".format(type(image)))
+                raise ValueError(f"Expected image input is a path or PIL.Image, but got {type(image)}.")
 
             if isinstance(image, str):
                 image = Image.open(image).convert("RGB")

src/llamafactory/cli.py

@@ -47,7 +47,7 @@ USAGE = (
 WELCOME = (
     "-" * 58
     + "\n"
-    + "| Welcome to LLaMA Factory, version {}".format(VERSION)
+    + f"| Welcome to LLaMA Factory, version {VERSION}"
     + " " * (21 - len(VERSION))
     + "|\n|"
     + " " * 56
@@ -90,7 +90,7 @@ def main():
         if force_torchrun or get_device_count() > 1:
             master_addr = os.environ.get("MASTER_ADDR", "127.0.0.1")
             master_port = os.environ.get("MASTER_PORT", str(random.randint(20001, 29999)))
-            logger.info("Initializing distributed tasks at: {}:{}".format(master_addr, master_port))
+            logger.info(f"Initializing distributed tasks at: {master_addr}:{master_port}")
             process = subprocess.run(
                 (
                     "torchrun --nnodes {nnodes} --node_rank {node_rank} --nproc_per_node {nproc_per_node} "
@@ -118,4 +118,4 @@ def main():
     elif command == Command.HELP:
         print(USAGE)
     else:
-        raise NotImplementedError("Unknown command: {}.".format(command))
+        raise NotImplementedError(f"Unknown command: {command}.")

src/llamafactory/data/aligner.py

@@ -161,7 +161,7 @@ def convert_sharegpt(
     broken_data = False
     for turn_idx, message in enumerate(messages):
         if message[dataset_attr.role_tag] not in accept_tags[turn_idx % 2]:
-            logger.warning("Invalid role tag in {}.".format(messages))
+            logger.warning(f"Invalid role tag in {messages}.")
             broken_data = True
 
         aligned_messages.append(
@@ -171,7 +171,7 @@ def convert_sharegpt(
     if (not dataset_attr.ranking and len(aligned_messages) % 2 != 0) or (
         dataset_attr.ranking and len(aligned_messages) % 2 == 0
     ):
-        logger.warning("Invalid message count in {}.".format(messages))
+        logger.warning(f"Invalid message count in {messages}.")
         broken_data = True
 
     if dataset_attr.kto_tag and isinstance(example[dataset_attr.kto_tag], bool):  # kto example
@@ -192,7 +192,7 @@ def convert_sharegpt(
             chosen[dataset_attr.role_tag] not in accept_tags[-1]
             or rejected[dataset_attr.role_tag] not in accept_tags[-1]
         ):
-            logger.warning("Invalid role tag in {}.".format([chosen, rejected]))
+            logger.warning(f"Invalid role tag in {[chosen, rejected]}.")
             broken_data = True
 
         prompt = aligned_messages

src/llamafactory/data/collator.py

@@ -137,9 +137,9 @@ class PairwiseDataCollatorWithPadding(MultiModalDataCollatorForSeq2Seq):
         for key in ("chosen", "rejected"):
             for feature in features:
                 target_feature = {
-                    "input_ids": feature["{}_input_ids".format(key)],
-                    "attention_mask": feature["{}_attention_mask".format(key)],
-                    "labels": feature["{}_labels".format(key)],
+                    "input_ids": feature[f"{key}_input_ids"],
+                    "attention_mask": feature[f"{key}_attention_mask"],
+                    "labels": feature[f"{key}_labels"],
                     "images": feature["images"],
                     "videos": feature["videos"],
                 }

src/llamafactory/data/data_utils.py

@@ -70,7 +70,7 @@ def merge_dataset(
             stopping_strategy="first_exhausted" if data_args.mix_strategy.endswith("under") else "all_exhausted",
         )
     else:
-        raise ValueError("Unknown mixing strategy: {}.".format(data_args.mix_strategy))
+        raise ValueError(f"Unknown mixing strategy: {data_args.mix_strategy}.")
 
 
 def split_dataset(

src/llamafactory/data/formatter.py

@@ -83,14 +83,14 @@ class StringFormatter(Formatter):
             if isinstance(slot, str):
                 for name, value in kwargs.items():
                     if not isinstance(value, str):
-                        raise RuntimeError("Expected a string, got {}".format(value))
+                        raise RuntimeError(f"Expected a string, got {value}")
 
                     slot = slot.replace("{{" + name + "}}", value, 1)
                 elements.append(slot)
             elif isinstance(slot, (dict, set)):
                 elements.append(slot)
             else:
-                raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
+                raise RuntimeError(f"Input must be string, set[str] or dict[str, str], got {type(slot)}")
 
         return elements
 
@@ -113,7 +113,7 @@ class FunctionFormatter(Formatter):
                 functions.append((tool_call["name"], json.dumps(tool_call["arguments"], ensure_ascii=False)))
 
         except json.JSONDecodeError:
-            raise RuntimeError("Invalid JSON format in function message: {}".format(str([content])))  # flat string
+            raise RuntimeError(f"Invalid JSON format in function message: {str([content])}")  # flat string
 
         elements = []
         for name, arguments in functions:
@@ -124,7 +124,7 @@ class FunctionFormatter(Formatter):
                 elif isinstance(slot, (dict, set)):
                     elements.append(slot)
                 else:
-                    raise RuntimeError("Input must be string, set[str] or dict[str, str], got {}".format(type(slot)))
+                    raise RuntimeError(f"Input must be string, set[str] or dict[str, str], got {type(slot)}")
 
         return elements
 
@@ -141,7 +141,7 @@ class ToolFormatter(Formatter):
             tools = json.loads(content)
             return [self.tool_utils.tool_formatter(tools) if len(tools) != 0 else ""]
         except json.JSONDecodeError:
-            raise RuntimeError("Invalid JSON format in tool description: {}".format(str([content])))  # flat string
+            raise RuntimeError(f"Invalid JSON format in tool description: {str([content])}")  # flat string
 
     @override
     def extract(self, content: str) -> Union[str, List["FunctionCall"]]:

src/llamafactory/data/loader.py

@@ -51,7 +51,7 @@ def _load_single_dataset(
     r"""
     Loads a single dataset and aligns it to the standard format.
     """
-    logger.info("Loading dataset {}...".format(dataset_attr))
+    logger.info(f"Loading dataset {dataset_attr}...")
     data_path, data_name, data_dir, data_files = None, None, None, None
     if dataset_attr.load_from in ["hf_hub", "ms_hub", "om_hub"]:
         data_path = dataset_attr.dataset_name
@@ -77,12 +77,12 @@ def _load_single_dataset(
             data_files.append(local_path)
             data_path = FILEEXT2TYPE.get(local_path.split(".")[-1], None)
         else:
-            raise ValueError("File {} not found.".format(local_path))
+            raise ValueError(f"File {local_path} not found.")
 
         if data_path is None:
             raise ValueError("Allowed file types: {}.".format(",".join(FILEEXT2TYPE.keys())))
     else:
-        raise NotImplementedError("Unknown load type: {}.".format(dataset_attr.load_from))
+        raise NotImplementedError(f"Unknown load type: {dataset_attr.load_from}.")
 
     if dataset_attr.load_from == "ms_hub":
         require_version("modelscope>=1.11.0", "To fix: pip install modelscope>=1.11.0")
@@ -145,7 +145,7 @@ def _load_single_dataset(
 
         assert len(indexes) == dataset_attr.num_samples, "Sample num mismatched."
         dataset = dataset.select(indexes)
-        logger.info("Sampled {} examples from dataset {}.".format(dataset_attr.num_samples, dataset_attr))
+        logger.info(f"Sampled {dataset_attr.num_samples} examples from dataset {dataset_attr}.")
 
     if data_args.max_samples is not None:  # truncate dataset
         max_samples = min(data_args.max_samples, len(dataset))
@@ -243,7 +243,7 @@ def get_dataset(
         if has_tokenized_data(data_args.tokenized_path):
             logger.warning("Loading dataset from disk will ignore other data arguments.")
             dataset_dict: "DatasetDict" = load_from_disk(data_args.tokenized_path)
-            logger.info("Loaded tokenized dataset from {}.".format(data_args.tokenized_path))
+            logger.info(f"Loaded tokenized dataset from {data_args.tokenized_path}.")
 
             dataset_module: Dict[str, "Dataset"] = {}
             if "train" in dataset_dict:
@@ -294,8 +294,8 @@ def get_dataset(
         if data_args.tokenized_path is not None:
             if training_args.should_save:
                 dataset_dict.save_to_disk(data_args.tokenized_path)
-                logger.info("Tokenized dataset saved at {}.".format(data_args.tokenized_path))
-                logger.info("Please restart the training with `tokenized_path: {}`.".format(data_args.tokenized_path))
+                logger.info(f"Tokenized dataset saved at {data_args.tokenized_path}.")
+                logger.info(f"Please restart the training with `tokenized_path: {data_args.tokenized_path}`.")
 
             sys.exit(0)
 

src/llamafactory/data/mm_plugin.py

@@ -111,7 +111,7 @@ class BasePlugin:
                     image = Image.open(image["path"])
 
             if not isinstance(image, ImageObject):
-                raise ValueError("Expect input is a list of Images, but got {}.".format(type(image)))
+                raise ValueError(f"Expect input is a list of Images, but got {type(image)}.")
 
             results.append(self._preprocess_image(image, **kwargs))
 
@@ -253,7 +253,7 @@ class LlavaPlugin(BasePlugin):
             message["content"] = content.replace("{{image}}", self.image_token * image_seqlen)
 
         if len(images) != num_image_tokens:
-            raise ValueError("The number of images does not match the number of {} tokens".format(IMAGE_PLACEHOLDER))
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens")
 
         return messages
 
@@ -302,7 +302,7 @@ class LlavaNextPlugin(BasePlugin):
             message["content"] = content.replace("{{image}}", self.image_token)
 
         if len(images) != num_image_tokens:
-            raise ValueError("The number of images does not match the number of {} tokens".format(IMAGE_PLACEHOLDER))
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens")
         return messages
 
     @override
@@ -366,10 +366,10 @@ class LlavaNextVideoPlugin(BasePlugin):
                 message["content"] = content.replace("{{video}}", self.video_token * video_seqlen)
 
         if len(images) != num_image_tokens:
-            raise ValueError("The number of images does not match the number of {} tokens".format(IMAGE_PLACEHOLDER))
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens")
 
         if len(videos) != num_video_tokens:
-            raise ValueError("The number of videos does not match the number of {} tokens".format(IMAGE_PLACEHOLDER))
+            raise ValueError(f"The number of videos does not match the number of {IMAGE_PLACEHOLDER} tokens")
 
         return messages
 
@@ -408,7 +408,7 @@ class PaliGemmaPlugin(BasePlugin):
             message["content"] = content.replace("{{image}}", "")
 
         if len(images) != num_image_tokens:
-            raise ValueError("The number of images does not match the number of {} tokens".format(IMAGE_PLACEHOLDER))
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens")
 
         return messages
 
@@ -493,7 +493,7 @@ class Qwen2vlPlugin(BasePlugin):
             content = message["content"]
             while IMAGE_PLACEHOLDER in content:
                 if num_image_tokens >= len(image_grid_thw):
-                    raise ValueError("`len(images)` is less than the number of {} tokens.".format(IMAGE_PLACEHOLDER))
+                    raise ValueError(f"`len(images)` is less than the number of {IMAGE_PLACEHOLDER} tokens.")
 
                 content = content.replace(
                     IMAGE_PLACEHOLDER,
@@ -506,7 +506,7 @@ class Qwen2vlPlugin(BasePlugin):
 
             while VIDEO_PLACEHOLDER in content:
                 if num_video_tokens >= len(video_grid_thw):
-                    raise ValueError("`len(videos)` is less than the number of {} tokens.".format(VIDEO_PLACEHOLDER))
+                    raise ValueError(f"`len(videos)` is less than the number of {VIDEO_PLACEHOLDER} tokens.")
 
                 content = content.replace(
                     VIDEO_PLACEHOLDER,
@@ -520,10 +520,10 @@ class Qwen2vlPlugin(BasePlugin):
             message["content"] = content
 
         if len(images) != num_image_tokens:
-            raise ValueError("The number of images does not match the number of {} tokens".format(IMAGE_PLACEHOLDER))
+            raise ValueError(f"The number of images does not match the number of {IMAGE_PLACEHOLDER} tokens")
 
         if len(videos) != num_video_tokens:
-            raise ValueError("The number of videos does not match the number of {} tokens".format(VIDEO_PLACEHOLDER))
+            raise ValueError(f"The number of videos does not match the number of {VIDEO_PLACEHOLDER} tokens")
 
         return messages
 
@@ -583,10 +583,10 @@ class VideoLlavaPlugin(BasePlugin):
                 message["content"] = content.replace("{{video}}", self.video_token * video_seqlen)
 
         if len(images) != num_image_tokens:
-            raise ValueError("The number of images does not match the number of {} tokens".format(self.image_token))
+            raise ValueError(f"The number of images does not match the number of {self.image_token} tokens")
 
         if len(videos) != num_video_tokens:
-            raise ValueError("The number of videos does not match the number of {} tokens".format(self.video_token))
+            raise ValueError(f"The number of videos does not match the number of {self.video_token} tokens")
 
         return messages
 
@@ -622,6 +622,6 @@ def get_mm_plugin(
 ) -> "BasePlugin":
     plugin_class = PLUGINS.get(name, None)
     if plugin_class is None:
-        raise ValueError("Multimodal plugin `{}` not found.".format(name))
+        raise ValueError(f"Multimodal plugin `{name}` not found.")
 
     return plugin_class(image_token, video_token)

src/llamafactory/data/parser.py

@@ -87,11 +87,11 @@ def get_dataset_list(dataset_names: Optional[Sequence[str]], dataset_dir: str) -
             config_path = os.path.join(dataset_dir, DATA_CONFIG)
 
         try:
-            with open(config_path, "r") as f:
+            with open(config_path) as f:
                 dataset_info = json.load(f)
         except Exception as err:
             if len(dataset_names) != 0:
-                raise ValueError("Cannot open {} due to {}.".format(config_path, str(err)))
+                raise ValueError(f"Cannot open {config_path} due to {str(err)}.")
 
             dataset_info = None
 
@@ -109,7 +109,7 @@ def get_dataset_list(dataset_names: Optional[Sequence[str]], dataset_dir: str) -
             continue
 
         if name not in dataset_info:
-            raise ValueError("Undefined dataset {} in {}.".format(name, DATA_CONFIG))
+            raise ValueError(f"Undefined dataset {name} in {DATA_CONFIG}.")
 
         has_hf_url = "hf_hub_url" in dataset_info[name]
         has_ms_url = "ms_hub_url" in dataset_info[name]

src/llamafactory/data/processors/pairwise.py

@@ -110,8 +110,8 @@ def print_pairwise_dataset_example(example: Dict[str, List[int]], tokenizer: "Pr
     print("chosen_input_ids:\n{}".format(example["chosen_input_ids"]))
     print("chosen_inputs:\n{}".format(tokenizer.decode(example["chosen_input_ids"], skip_special_tokens=False)))
     print("chosen_label_ids:\n{}".format(example["chosen_labels"]))
-    print("chosen_labels:\n{}".format(tokenizer.decode(valid_chosen_labels, skip_special_tokens=False)))
+    print(f"chosen_labels:\n{tokenizer.decode(valid_chosen_labels, skip_special_tokens=False)}")
     print("rejected_input_ids:\n{}".format(example["rejected_input_ids"]))
     print("rejected_inputs:\n{}".format(tokenizer.decode(example["rejected_input_ids"], skip_special_tokens=False)))
     print("rejected_label_ids:\n{}".format(example["rejected_labels"]))
-    print("rejected_labels:\n{}".format(tokenizer.decode(valid_rejected_labels, skip_special_tokens=False)))
+    print(f"rejected_labels:\n{tokenizer.decode(valid_rejected_labels, skip_special_tokens=False)}")

src/llamafactory/data/processors/supervised.py

@@ -160,7 +160,7 @@ def preprocess_packed_supervised_dataset(
         )
         length = len(input_ids)
         if length > data_args.cutoff_len:
-            logger.warning("Dropped lengthy example with length {} > {}.".format(length, data_args.cutoff_len))
+            logger.warning(f"Dropped lengthy example with length {length} > {data_args.cutoff_len}.")
         else:
             lengths.append(length)
             length2indexes[length].append(valid_num)
@@ -212,4 +212,4 @@ def print_supervised_dataset_example(example: Dict[str, List[int]], tokenizer: "
     print("input_ids:\n{}".format(example["input_ids"]))
     print("inputs:\n{}".format(tokenizer.decode(example["input_ids"], skip_special_tokens=False)))
     print("label_ids:\n{}".format(example["labels"]))
-    print("labels:\n{}".format(tokenizer.decode(valid_labels, skip_special_tokens=False)))
+    print(f"labels:\n{tokenizer.decode(valid_labels, skip_special_tokens=False)}")

src/llamafactory/data/template.py

@@ -147,7 +147,7 @@ class Template:
                 elif "eos_token" in elem and tokenizer.eos_token_id is not None:
                     token_ids += [tokenizer.eos_token_id]
             else:
-                raise ValueError("Input must be string, set[str] or dict[str, str], got {}".format(type(elem)))
+                raise ValueError(f"Input must be string, set[str] or dict[str, str], got {type(elem)}")
 
         return token_ids
 
@@ -275,9 +275,9 @@ def _add_or_replace_eos_token(tokenizer: "PreTrainedTokenizer", eos_token: str)
     num_added_tokens = tokenizer.add_special_tokens({"eos_token": eos_token})
 
     if is_added:
-        logger.info("Add eos token: {}".format(tokenizer.eos_token))
+        logger.info(f"Add eos token: {tokenizer.eos_token}")
     else:
-        logger.info("Replace eos token: {}".format(tokenizer.eos_token))
+        logger.info(f"Replace eos token: {tokenizer.eos_token}")
 
     if num_added_tokens > 0:
         logger.warning("New tokens have been added, make sure `resize_vocab` is True.")
@@ -365,13 +365,13 @@ def get_template_and_fix_tokenizer(tokenizer: "PreTrainedTokenizer", data_args:
     else:
         template = TEMPLATES.get(data_args.template, None)
         if template is None:
-            raise ValueError("Template {} does not exist.".format(data_args.template))
+            raise ValueError(f"Template {data_args.template} does not exist.")
 
     if data_args.train_on_prompt and template.efficient_eos:
         raise ValueError("Current template does not support `train_on_prompt`.")
 
     if data_args.tool_format is not None:
-        logger.info("Using tool format: {}.".format(data_args.tool_format))
+        logger.info(f"Using tool format: {data_args.tool_format}.")
         eos_slots = [] if template.efficient_eos else [{"eos_token"}]
         template.format_function = FunctionFormatter(slots=eos_slots, tool_format=data_args.tool_format)
         template.format_tools = ToolFormatter(tool_format=data_args.tool_format)
@@ -389,7 +389,7 @@ def get_template_and_fix_tokenizer(tokenizer: "PreTrainedTokenizer", data_args:
 
     if tokenizer.pad_token_id is None:
         tokenizer.pad_token = tokenizer.eos_token
-        logger.info("Add pad token: {}".format(tokenizer.pad_token))
+        logger.info(f"Add pad token: {tokenizer.pad_token}")
 
     if stop_words:
         num_added_tokens = tokenizer.add_special_tokens(

src/llamafactory/data/tool_utils.py

@@ -177,6 +177,6 @@ TOOLS = {
 def get_tool_utils(name: str) -> "ToolUtils":
     tool_utils = TOOLS.get(name, None)
     if tool_utils is None:
-        raise ValueError("Tool utils `{}` not found.".format(name))
+        raise ValueError(f"Tool utils `{name}` not found.")
 
     return tool_utils

src/llamafactory/eval/evaluator.py

@@ -87,7 +87,7 @@ class Evaluator:
             token=self.model_args.hf_hub_token,
         )
 
-        with open(mapping, "r", encoding="utf-8") as f:
+        with open(mapping, encoding="utf-8") as f:
             categorys: Dict[str, Dict[str, str]] = json.load(f)
 
         category_corrects = {subj: np.array([], dtype="bool") for subj in SUBJECTS}
@@ -139,7 +139,7 @@ class Evaluator:
     def _save_results(self, category_corrects: Dict[str, "NDArray"], results: Dict[str, Dict[int, str]]) -> None:
         score_info = "\n".join(
             [
-                "{:>15}: {:.2f}".format(category_name, 100 * np.mean(category_correct))
+                f"{category_name:>15}: {100 * np.mean(category_correct):.2f}"
                 for category_name, category_correct in category_corrects.items()
                 if len(category_correct)
             ]

src/llamafactory/eval/template.py

@@ -61,7 +61,7 @@ def _register_eval_template(name: str, system: str, choice: str, answer: str) ->
 
 def get_eval_template(name: str) -> "EvalTemplate":
     eval_template = eval_templates.get(name, None)
-    assert eval_template is not None, "Template {} does not exist.".format(name)
+    assert eval_template is not None, f"Template {name} does not exist."
     return eval_template
 
 

src/llamafactory/extras/env.py

@@ -72,4 +72,4 @@ def print_env() -> None:
     except Exception:
         pass
 
-    print("\n" + "\n".join(["- {}: {}".format(key, value) for key, value in info.items()]) + "\n")
+    print("\n" + "\n".join([f"- {key}: {value}" for key, value in info.items()]) + "\n")

src/llamafactory/extras/logging.py

@@ -75,7 +75,7 @@ def _get_default_logging_level() -> "logging._Level":
         if env_level_str.upper() in logging._nameToLevel:
             return logging._nameToLevel[env_level_str.upper()]
         else:
-            raise ValueError("Unknown logging level: {}.".format(env_level_str))
+            raise ValueError(f"Unknown logging level: {env_level_str}.")
 
     return _default_log_level
 

src/llamafactory/extras/ploting.py

@@ -75,7 +75,7 @@ def plot_loss(save_dictionary: str, keys: List[str] = ["loss"]) -> None:
     Plots loss curves and saves the image.
     """
     plt.switch_backend("agg")
-    with open(os.path.join(save_dictionary, TRAINER_STATE_NAME), "r", encoding="utf-8") as f:
+    with open(os.path.join(save_dictionary, TRAINER_STATE_NAME), encoding="utf-8") as f:
         data = json.load(f)
 
     for key in keys:
@@ -92,7 +92,7 @@ def plot_loss(save_dictionary: str, keys: List[str] = ["loss"]) -> None:
         plt.figure()
         plt.plot(steps, metrics, color="#1f77b4", alpha=0.4, label="original")
         plt.plot(steps, smooth(metrics), color="#1f77b4", label="smoothed")
-        plt.title("training {} of {}".format(key, save_dictionary))
+        plt.title(f"training {key} of {save_dictionary}")
         plt.xlabel("step")
         plt.ylabel(key)
         plt.legend()

src/llamafactory/hparams/parser.py

@@ -67,8 +67,8 @@ def _parse_args(parser: "HfArgumentParser", args: Optional[Dict[str, Any]] = Non
 
     if unknown_args:
         print(parser.format_help())
-        print("Got unknown args, potentially deprecated arguments: {}".format(unknown_args))
-        raise ValueError("Some specified arguments are not used by the HfArgumentParser: {}".format(unknown_args))
+        print(f"Got unknown args, potentially deprecated arguments: {unknown_args}")
+        raise ValueError(f"Some specified arguments are not used by the HfArgumentParser: {unknown_args}")
 
     return (*parsed_args,)
 
@@ -323,7 +323,7 @@ def get_train_args(args: Optional[Dict[str, Any]] = None) -> _TRAIN_CLS:
 
         if last_checkpoint is not None:
             training_args.resume_from_checkpoint = last_checkpoint
-            logger.info("Resuming training from {}.".format(training_args.resume_from_checkpoint))
+            logger.info(f"Resuming training from {training_args.resume_from_checkpoint}.")
             logger.info("Change `output_dir` or use `overwrite_output_dir` to avoid.")
 
     if (

src/llamafactory/model/adapter.py

@@ -182,7 +182,7 @@ def _setup_lora_tuning(
             model = model.merge_and_unload()
 
         if len(adapter_to_merge) > 0:
-            logger.info("Merged {} adapter(s).".format(len(adapter_to_merge)))
+            logger.info(f"Merged {len(adapter_to_merge)} adapter(s).")
 
         if adapter_to_resume is not None:  # resume lora training
             if model_args.use_unsloth:
@@ -239,8 +239,8 @@ def _setup_lora_tuning(
                     logger.info("Using PiSSA initialization.")
                     peft_kwargs["init_lora_weights"] = "pissa"
                 else:
-                    logger.info("Using PiSSA initialization with FSVD steps {}.".format(finetuning_args.pissa_iter))
-                    peft_kwargs["init_lora_weights"] = "pissa_niter_{}".format(finetuning_args.pissa_iter)
+                    logger.info(f"Using PiSSA initialization with FSVD steps {finetuning_args.pissa_iter}.")
+                    peft_kwargs["init_lora_weights"] = f"pissa_niter_{finetuning_args.pissa_iter}"
 
             lora_config = LoraConfig(
                 task_type=TaskType.CAUSAL_LM,
@@ -300,6 +300,6 @@ def init_adapter(
             config, model, model_args, finetuning_args, is_trainable, cast_trainable_params_to_fp32
         )
     else:
-        raise NotImplementedError("Unknown finetuning type: {}.".format(finetuning_args.finetuning_type))
+        raise NotImplementedError(f"Unknown finetuning type: {finetuning_args.finetuning_type}.")
 
     return model

src/llamafactory/model/loader.py

@@ -100,7 +100,7 @@ def load_tokenizer(model_args: "ModelArguments") -> "TokenizerModule":
         processor = AutoProcessor.from_pretrained(model_args.model_name_or_path, **init_kwargs)
         patch_processor(processor, config, tokenizer, model_args)
     except Exception as e:
-        logger.warning("Processor was not found: {}.".format(e))
+        logger.warning(f"Processor was not found: {e}.")
         processor = None
 
     # Avoid load tokenizer, see:
@@ -180,7 +180,7 @@ def load_model(
         vhead_params = load_valuehead_params(vhead_path, model_args)
         if vhead_params is not None:
             model.load_state_dict(vhead_params, strict=False)
-            logger.info("Loaded valuehead from checkpoint: {}".format(vhead_path))
+            logger.info(f"Loaded valuehead from checkpoint: {vhead_path}")
 
     if not is_trainable:
         model.requires_grad_(False)
@@ -198,7 +198,7 @@ def load_model(
             trainable_params, all_param, 100 * trainable_params / all_param
         )
     else:
-        param_stats = "all params: {:,}".format(all_param)
+        param_stats = f"all params: {all_param:,}"
 
     logger.info(param_stats)
 

src/llamafactory/model/model_utils/attention.py

@@ -65,7 +65,7 @@ def configure_attn_implementation(
 
         requested_attn_implementation = "flash_attention_2"
     else:
-        raise NotImplementedError("Unknown attention type: {}".format(model_args.flash_attn))
+        raise NotImplementedError(f"Unknown attention type: {model_args.flash_attn}")
 
     if getattr(config, "model_type", None) == "internlm2":  # special case for custom models
         setattr(config, "attn_implementation", requested_attn_implementation)

src/llamafactory/model/model_utils/checkpointing.py

@@ -111,7 +111,7 @@ def _gradient_checkpointing_enable(
     from torch.utils.checkpoint import checkpoint
 
     if not self.supports_gradient_checkpointing:
-        raise ValueError("{} does not support gradient checkpointing.".format(self.__class__.__name__))
+        raise ValueError(f"{self.__class__.__name__} does not support gradient checkpointing.")
 
     if gradient_checkpointing_kwargs is None:
         gradient_checkpointing_kwargs = {"use_reentrant": True}

src/llamafactory/model/model_utils/embedding.py

@@ -69,4 +69,4 @@ def resize_embedding_layer(model: "PreTrainedModel", tokenizer: "PreTrainedToken
             _noisy_mean_initialization(model.get_input_embeddings().weight.data, num_new_tokens)
             _noisy_mean_initialization(model.get_output_embeddings().weight.data, num_new_tokens)
 
-        logger.info("Resized token embeddings from {} to {}.".format(current_embedding_size, new_embedding_size))
+        logger.info(f"Resized token embeddings from {current_embedding_size} to {new_embedding_size}.")

src/llamafactory/model/model_utils/misc.py

@@ -67,12 +67,12 @@ def find_expanded_modules(model: "PreTrainedModel", target_modules: List[str], n
 
     if num_layers % num_layer_trainable != 0:
         raise ValueError(
-            "`num_layers` {} should be divisible by `num_layer_trainable` {}.".format(num_layers, num_layer_trainable)
+            f"`num_layers` {num_layers} should be divisible by `num_layer_trainable` {num_layer_trainable}."
         )
 
     stride = num_layers // num_layer_trainable
     trainable_layer_ids = range(stride - 1, num_layers + stride - 1, stride)
-    trainable_layers = [".{:d}.".format(idx) for idx in trainable_layer_ids]
+    trainable_layers = [f".{idx:d}." for idx in trainable_layer_ids]
     module_names = []
     for name, _ in model.named_modules():
         if any(target_module in name for target_module in target_modules) and any(

src/llamafactory/model/model_utils/quantization.py

@@ -130,7 +130,7 @@ def configure_quantization(
             quantization_config["bits"] = 2
 
         quant_bits = quantization_config.get("bits", "?")
-        logger.info("Loading {}-bit {}-quantized model.".format(quant_bits, quant_method.upper()))
+        logger.info(f"Loading {quant_bits}-bit {quant_method.upper()}-quantized model.")
 
     elif model_args.export_quantization_bit is not None:  # auto-gptq
         if model_args.export_quantization_bit not in [8, 4, 3, 2]:
@@ -149,7 +149,7 @@ def configure_quantization(
         )
         init_kwargs["device_map"] = "auto"
         init_kwargs["max_memory"] = get_max_memory()
-        logger.info("Quantizing model to {} bit with AutoGPTQ.".format(model_args.export_quantization_bit))
+        logger.info(f"Quantizing model to {model_args.export_quantization_bit} bit with AutoGPTQ.")
 
     elif model_args.quantization_bit is not None:  # on-the-fly
         if model_args.quantization_method == QuantizationMethod.BITS_AND_BYTES.value:
@@ -179,7 +179,7 @@ def configure_quantization(
             else:
                 init_kwargs["device_map"] = {"": get_current_device()}  # change auto device map for inference
 
-            logger.info("Quantizing model to {} bit with bitsandbytes.".format(model_args.quantization_bit))
+            logger.info(f"Quantizing model to {model_args.quantization_bit} bit with bitsandbytes.")
         elif model_args.quantization_method == QuantizationMethod.HQQ.value:
             if model_args.quantization_bit not in [8, 6, 5, 4, 3, 2, 1]:
                 raise ValueError("HQQ only accepts 1/2/3/4/5/6/8-bit quantization.")
@@ -191,7 +191,7 @@ def configure_quantization(
             init_kwargs["quantization_config"] = HqqConfig(
                 nbits=model_args.quantization_bit, quant_zero=False, quant_scale=False, axis=0
             )  # use ATEN kernel (axis=0) for performance
-            logger.info("Quantizing model to {} bit with HQQ.".format(model_args.quantization_bit))
+            logger.info(f"Quantizing model to {model_args.quantization_bit} bit with HQQ.")
         elif model_args.quantization_method == QuantizationMethod.EETQ.value:
             if model_args.quantization_bit != 8:
                 raise ValueError("EETQ only accepts 8-bit quantization.")
@@ -201,4 +201,4 @@ def configure_quantization(
 
             require_version("eetq", "To fix: pip install eetq")
             init_kwargs["quantization_config"] = EetqConfig()
-            logger.info("Quantizing model to {} bit with EETQ.".format(model_args.quantization_bit))
+            logger.info(f"Quantizing model to {model_args.quantization_bit} bit with EETQ.")

src/llamafactory/model/model_utils/rope.py

@@ -48,9 +48,7 @@ def configure_rope(config: "PretrainedConfig", model_args: "ModelArguments", is_
 
         current_max_length = getattr(config, "max_position_embeddings", None)
         if current_max_length and model_args.model_max_length > current_max_length:
-            logger.info(
-                "Enlarge max model length from {} to {}.".format(current_max_length, model_args.model_max_length)
-            )
+            logger.info(f"Enlarge max model length from {current_max_length} to {model_args.model_max_length}.")
             setattr(config, "max_position_embeddings", model_args.model_max_length)
             scaling_factor = float(math.ceil(model_args.model_max_length / current_max_length))
         else:
@@ -60,6 +58,4 @@ def configure_rope(config: "PretrainedConfig", model_args: "ModelArguments", is_
         scaling_factor = 2.0
 
     setattr(config, "rope_scaling", {"type": model_args.rope_scaling, "factor": scaling_factor})
-    logger.info(
-        "Using {} scaling strategy and setting scaling factor to {}".format(model_args.rope_scaling, scaling_factor)
-    )
+    logger.info(f"Using {model_args.rope_scaling} scaling strategy and setting scaling factor to {scaling_factor}")

src/llamafactory/model/model_utils/valuehead.py

@@ -54,7 +54,7 @@ def load_valuehead_params(path_or_repo_id: str, model_args: "ModelArguments") ->
     except Exception as err:
         err_text = str(err)
 
-    logger.info("Provided path ({}) does not contain value head weights: {}.".format(path_or_repo_id, err_text))
+    logger.info(f"Provided path ({path_or_repo_id}) does not contain value head weights: {err_text}.")
     logger.info("Ignore the above message if you are not resuming the training of a value head model.")
     return None
 

src/llamafactory/model/model_utils/visual.py

@@ -99,7 +99,7 @@ def autocast_projector_dtype(model: "PreTrainedModel", model_args: "ModelArgumen
         else:
             return
 
-        logger.info("Casting multimodal projector outputs in {}.".format(model_args.compute_dtype))
+        logger.info(f"Casting multimodal projector outputs in {model_args.compute_dtype}.")
         mm_projector.register_forward_hook(_mm_projector_forward_post_hook)
 
 

src/llamafactory/train/callbacks.py

@@ -92,7 +92,7 @@ def fix_valuehead_checkpoint(
     else:
         torch.save(v_head_state_dict, os.path.join(output_dir, V_HEAD_WEIGHTS_NAME))
 
-    logger.info("Value head model saved at: {}".format(output_dir))
+    logger.info(f"Value head model saved at: {output_dir}")
 
 
 class FixValueHeadModelCallback(TrainerCallback):
@@ -106,7 +106,7 @@ class FixValueHeadModelCallback(TrainerCallback):
         Event called after a checkpoint save.
         """
         if args.should_save:
-            output_dir = os.path.join(args.output_dir, "{}-{}".format(PREFIX_CHECKPOINT_DIR, state.global_step))
+            output_dir = os.path.join(args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}")
             fix_valuehead_checkpoint(
                 model=kwargs.pop("model"), output_dir=output_dir, safe_serialization=args.save_safetensors
             )
@@ -123,7 +123,7 @@ class SaveProcessorCallback(TrainerCallback):
     @override
     def on_save(self, args: "TrainingArguments", state: "TrainerState", control: "TrainerControl", **kwargs):
         if args.should_save:
-            output_dir = os.path.join(args.output_dir, "{}-{}".format(PREFIX_CHECKPOINT_DIR, state.global_step))
+            output_dir = os.path.join(args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{state.global_step}")
             getattr(self.processor, "image_processor").save_pretrained(output_dir)
 
     @override
@@ -145,7 +145,7 @@ class PissaConvertCallback(TrainerCallback):
         if args.should_save:
             model = kwargs.pop("model")
             pissa_init_dir = os.path.join(args.output_dir, "pissa_init")
-            logger.info("Initial PiSSA adapter will be saved at: {}.".format(pissa_init_dir))
+            logger.info(f"Initial PiSSA adapter will be saved at: {pissa_init_dir}.")
             if isinstance(model, PeftModel):
                 init_lora_weights = getattr(model.peft_config["default"], "init_lora_weights")
                 setattr(model.peft_config["default"], "init_lora_weights", True)
@@ -159,7 +159,7 @@ class PissaConvertCallback(TrainerCallback):
             pissa_init_dir = os.path.join(args.output_dir, "pissa_init")
             pissa_backup_dir = os.path.join(args.output_dir, "pissa_backup")
             pissa_convert_dir = os.path.join(args.output_dir, "pissa_converted")
-            logger.info("Converted PiSSA adapter will be saved at: {}.".format(pissa_convert_dir))
+            logger.info(f"Converted PiSSA adapter will be saved at: {pissa_convert_dir}.")
             # 1. save a pissa backup with init_lora_weights: True
             # 2. save a converted lora with init_lora_weights: pissa
             # 3. load the pissa backup with init_lora_weights: True

src/llamafactory/train/dpo/trainer.py

@@ -156,7 +156,7 @@ class CustomDPOTrainer(DPOTrainer):
             elif self.loss_type == "simpo":
                 losses = self.simpo_loss(policy_chosen_logps, policy_rejected_logps)
             else:
-                raise NotImplementedError("Unknown loss type: {}.".format(self.loss_type))
+                raise NotImplementedError(f"Unknown loss type: {self.loss_type}.")
 
             chosen_rewards = self.beta * policy_chosen_logps.to(self.accelerator.device).detach()
             rejected_rewards = self.beta * policy_rejected_logps.to(self.accelerator.device).detach()
@@ -245,16 +245,16 @@ class CustomDPOTrainer(DPOTrainer):
         reward_accuracies = (chosen_rewards > rejected_rewards).float()
 
         prefix = "eval_" if train_eval == "eval" else ""
-        metrics["{}rewards/chosen".format(prefix)] = chosen_rewards.mean().cpu()
-        metrics["{}rewards/rejected".format(prefix)] = rejected_rewards.mean().cpu()
-        metrics["{}rewards/accuracies".format(prefix)] = reward_accuracies.mean().cpu()
-        metrics["{}rewards/margins".format(prefix)] = (chosen_rewards - rejected_rewards).mean().cpu()
-        metrics["{}logps/rejected".format(prefix)] = policy_rejected_logps.detach().mean().cpu()
-        metrics["{}logps/chosen".format(prefix)] = policy_chosen_logps.detach().mean().cpu()
-        metrics["{}logits/rejected".format(prefix)] = policy_rejected_logits.detach().mean().cpu()
-        metrics["{}logits/chosen".format(prefix)] = policy_chosen_logits.detach().mean().cpu()
+        metrics[f"{prefix}rewards/chosen"] = chosen_rewards.mean().cpu()
+        metrics[f"{prefix}rewards/rejected"] = rejected_rewards.mean().cpu()
+        metrics[f"{prefix}rewards/accuracies"] = reward_accuracies.mean().cpu()
+        metrics[f"{prefix}rewards/margins"] = (chosen_rewards - rejected_rewards).mean().cpu()
+        metrics[f"{prefix}logps/rejected"] = policy_rejected_logps.detach().mean().cpu()
+        metrics[f"{prefix}logps/chosen"] = policy_chosen_logps.detach().mean().cpu()
+        metrics[f"{prefix}logits/rejected"] = policy_rejected_logits.detach().mean().cpu()
+        metrics[f"{prefix}logits/chosen"] = policy_chosen_logits.detach().mean().cpu()
         if self.loss_type == "orpo":
-            metrics["{}sft_loss".format(prefix)] = sft_loss.detach().mean().cpu()
-            metrics["{}odds_ratio_loss".format(prefix)] = ((losses - sft_loss) / self.beta).detach().mean().cpu()
+            metrics[f"{prefix}sft_loss"] = sft_loss.detach().mean().cpu()
+            metrics[f"{prefix}odds_ratio_loss"] = ((losses - sft_loss) / self.beta).detach().mean().cpu()
 
         return losses.mean(), metrics

src/llamafactory/train/kto/trainer.py

@@ -129,11 +129,11 @@ class CustomKTOTrainer(KTOTrainer):
         """
         batch = {k: v.detach().clone() for k, v in batch.items()}  # avoid error
         model_inputs = {
-            "input_ids": batch["{}input_ids".format(prefix)],
-            "attention_mask": batch["{}attention_mask".format(prefix)],
+            "input_ids": batch[f"{prefix}input_ids"],
+            "attention_mask": batch[f"{prefix}attention_mask"],
         }
-        if "{}token_type_ids".format(prefix) in batch:
-            model_inputs["token_type_ids"] = batch["{}token_type_ids".format(prefix)]
+        if f"{prefix}token_type_ids" in batch:
+            model_inputs["token_type_ids"] = batch[f"{prefix}token_type_ids"]
 
         if "pixel_values" in batch:
             model_inputs["pixel_values"] = batch["pixel_values"]
@@ -142,7 +142,7 @@ class CustomKTOTrainer(KTOTrainer):
             model_inputs["image_grid_thw"] = batch["image_grid_thw"]
 
         logits = model(**model_inputs, return_dict=True, use_cache=False).logits.to(torch.float32)
-        logps, valid_length = get_batch_logps(logits=logits, labels=batch["{}labels".format(prefix)])
+        logps, valid_length = get_batch_logps(logits=logits, labels=batch[f"{prefix}labels"])
         return logps, logps / valid_length
 
     @override

src/llamafactory/train/ppo/ppo_utils.py

@@ -62,8 +62,8 @@ def replace_model(model: "AutoModelForCausalLMWithValueHead", target: Literal["d
             setattr(model, "default_head_bias", v_head_layer.bias.data.detach().clone())
 
         device = v_head_layer.weight.device
-        v_head_layer.weight.data = model.get_buffer("{}_head_weight".format(target)).detach().clone().to(device)
-        v_head_layer.bias.data = model.get_buffer("{}_head_bias".format(target)).detach().clone().to(device)
+        v_head_layer.weight.data = model.get_buffer(f"{target}_head_weight").detach().clone().to(device)
+        v_head_layer.bias.data = model.get_buffer(f"{target}_head_bias").detach().clone().to(device)
 
 
 def dump_layernorm(model: "PreTrainedModel") -> Dict[str, "torch.Tensor"]:

src/llamafactory/train/ppo/trainer.py

@@ -218,18 +218,18 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
 
         if self.is_world_process_zero():
             logger.info("***** Running training *****")
-            logger.info("  Num examples = {:,}".format(num_examples))
-            logger.info("  Num Epochs = {:,}".format(num_train_epochs))
-            logger.info("  Instantaneous batch size per device = {:,}".format(self.args.per_device_train_batch_size))
+            logger.info(f"  Num examples = {num_examples:,}")
+            logger.info(f"  Num Epochs = {num_train_epochs:,}")
+            logger.info(f"  Instantaneous batch size per device = {self.args.per_device_train_batch_size:,}")
             logger.info(
                 "  Total train batch size (w. parallel, buffer, distributed & accumulation) = {:,}".format(
                     total_train_batch_size
                 )
             )
-            logger.info("  Gradient Accumulation steps = {:,}".format(self.args.gradient_accumulation_steps))
-            logger.info("  Num optimization epochs per batch = {:,}".format(self.finetuning_args.ppo_epochs))
-            logger.info("  Total training steps = {:,}".format(max_steps))
-            logger.info("  Number of trainable parameters = {:,}".format(count_parameters(self.model)[0]))
+            logger.info(f"  Gradient Accumulation steps = {self.args.gradient_accumulation_steps:,}")
+            logger.info(f"  Num optimization epochs per batch = {self.finetuning_args.ppo_epochs:,}")
+            logger.info(f"  Total training steps = {max_steps:,}")
+            logger.info(f"  Number of trainable parameters = {count_parameters(self.model)[0]:,}")
 
         dataiter = iter(self.dataloader)
         loss_meter = AverageMeter()
@@ -290,7 +290,7 @@ class CustomPPOTrainer(PPOTrainer, Trainer):
 
             if (step + 1) % self.args.save_steps == 0:  # save checkpoint
                 self.save_model(
-                    os.path.join(self.args.output_dir, "{}-{}".format(PREFIX_CHECKPOINT_DIR, self.state.global_step))
+                    os.path.join(self.args.output_dir, f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}")
                 )
                 self.callback_handler.on_save(self.args, self.state, self.control)
 

src/llamafactory/train/trainer_utils.py

@@ -116,7 +116,7 @@ def create_ref_model(
         ref_model = load_model(
             tokenizer, ref_model_args, ref_finetuning_args, is_trainable=False, add_valuehead=add_valuehead
         )
-        logger.info("Created reference model from {}".format(finetuning_args.ref_model))
+        logger.info(f"Created reference model from {finetuning_args.ref_model}")
     else:
         if finetuning_args.finetuning_type == "lora":
             ref_model = None
@@ -140,7 +140,7 @@ def create_reward_model(
     """
     if finetuning_args.reward_model_type == "api":
         assert finetuning_args.reward_model.startswith("http"), "Please provide full url."
-        logger.info("Use reward server {}".format(finetuning_args.reward_model))
+        logger.info(f"Use reward server {finetuning_args.reward_model}")
         return finetuning_args.reward_model
     elif finetuning_args.reward_model_type == "lora":
         model.pretrained_model.load_adapter(finetuning_args.reward_model, "reward")
@@ -157,7 +157,7 @@ def create_reward_model(
         model.register_buffer(
             "default_head_bias", torch.zeros_like(vhead_params["v_head.summary.bias"]), persistent=False
         )
-        logger.info("Loaded adapter weights of reward model from {}".format(finetuning_args.reward_model))
+        logger.info(f"Loaded adapter weights of reward model from {finetuning_args.reward_model}")
         return None
     else:
         reward_model_args = ModelArguments.copyfrom(
@@ -171,7 +171,7 @@ def create_reward_model(
         reward_model = load_model(
             tokenizer, reward_model_args, reward_finetuning_args, is_trainable=False, add_valuehead=True
         )
-        logger.info("Loaded full weights of reward model from {}".format(finetuning_args.reward_model))
+        logger.info(f"Loaded full weights of reward model from {finetuning_args.reward_model}")
         logger.warning("Please ensure the ppo model and reward model share SAME tokenizer and vocabulary.")
         return reward_model
 
@@ -231,7 +231,7 @@ def _create_galore_optimizer(
     elif training_args.optim == "adafactor":
         optim_class = GaLoreAdafactor
     else:
-        raise NotImplementedError("Unknow optim: {}".format(training_args.optim))
+        raise NotImplementedError(f"Unknow optim: {training_args.optim}")
 
     if finetuning_args.galore_layerwise:
         if training_args.gradient_accumulation_steps != 1:
@@ -305,7 +305,7 @@ def _create_loraplus_optimizer(
         dict(params=param_dict["embedding"], lr=embedding_lr, weight_decay=training_args.weight_decay),
     ]
     optimizer = optim_class(param_groups, **optim_kwargs)
-    logger.info("Using LoRA+ optimizer with loraplus lr ratio {:.2f}.".format(finetuning_args.loraplus_lr_ratio))
+    logger.info(f"Using LoRA+ optimizer with loraplus lr ratio {finetuning_args.loraplus_lr_ratio:.2f}.")
     return optimizer
 
 

src/llamafactory/train/tuner.py

@@ -57,7 +57,7 @@ def run_exp(args: Optional[Dict[str, Any]] = None, callbacks: List["TrainerCallb
     elif finetuning_args.stage == "kto":
         run_kto(model_args, data_args, training_args, finetuning_args, callbacks)
     else:
-        raise ValueError("Unknown task: {}.".format(finetuning_args.stage))
+        raise ValueError(f"Unknown task: {finetuning_args.stage}.")
 
 
 def export_model(args: Optional[Dict[str, Any]] = None) -> None:
@@ -91,18 +91,18 @@ def export_model(args: Optional[Dict[str, Any]] = None) -> None:
 
         setattr(model.config, "torch_dtype", output_dtype)
         model = model.to(output_dtype)
-        logger.info("Convert model dtype to: {}.".format(output_dtype))
+        logger.info(f"Convert model dtype to: {output_dtype}.")
 
     model.save_pretrained(
         save_directory=model_args.export_dir,
-        max_shard_size="{}GB".format(model_args.export_size),
+        max_shard_size=f"{model_args.export_size}GB",
         safe_serialization=(not model_args.export_legacy_format),
     )
     if model_args.export_hub_model_id is not None:
         model.push_to_hub(
             model_args.export_hub_model_id,
             token=model_args.hf_hub_token,
-            max_shard_size="{}GB".format(model_args.export_size),
+            max_shard_size=f"{model_args.export_size}GB",
             safe_serialization=(not model_args.export_legacy_format),
         )
 
@@ -117,13 +117,13 @@ def export_model(args: Optional[Dict[str, Any]] = None) -> None:
                 os.path.join(vhead_path, V_HEAD_SAFE_WEIGHTS_NAME),
                 os.path.join(model_args.export_dir, V_HEAD_SAFE_WEIGHTS_NAME),
             )
-            logger.info("Copied valuehead to {}.".format(model_args.export_dir))
+            logger.info(f"Copied valuehead to {model_args.export_dir}.")
         elif os.path.exists(os.path.join(vhead_path, V_HEAD_WEIGHTS_NAME)):
             shutil.copy(
                 os.path.join(vhead_path, V_HEAD_WEIGHTS_NAME),
                 os.path.join(model_args.export_dir, V_HEAD_WEIGHTS_NAME),
             )
-            logger.info("Copied valuehead to {}.".format(model_args.export_dir))
+            logger.info(f"Copied valuehead to {model_args.export_dir}.")
 
     try:
         tokenizer.padding_side = "left"  # restore padding side
@@ -140,4 +140,4 @@ def export_model(args: Optional[Dict[str, Any]] = None) -> None:
                 )
 
     except Exception as e:
-        logger.warning("Cannot save tokenizer, please copy the files manually: {}.".format(e))
+        logger.warning(f"Cannot save tokenizer, please copy the files manually: {e}.")

src/llamafactory/webui/common.py

@@ -75,7 +75,7 @@ def load_config() -> Dict[str, Any]:
     Loads user config if exists.
     """
     try:
-        with open(get_config_path(), "r", encoding="utf-8") as f:
+        with open(get_config_path(), encoding="utf-8") as f:
             return safe_load(f)
     except Exception:
         return {"lang": None, "last_model": None, "path_dict": {}, "cache_dir": None}
@@ -172,14 +172,14 @@ def load_dataset_info(dataset_dir: str) -> Dict[str, Dict[str, Any]]:
     Loads dataset_info.json.
     """
     if dataset_dir == "ONLINE" or dataset_dir.startswith("REMOTE:"):
-        logger.info("dataset_dir is {}, using online dataset.".format(dataset_dir))
+        logger.info(f"dataset_dir is {dataset_dir}, using online dataset.")
         return {}
 
     try:
-        with open(os.path.join(dataset_dir, DATA_CONFIG), "r", encoding="utf-8") as f:
+        with open(os.path.join(dataset_dir, DATA_CONFIG), encoding="utf-8") as f:
             return json.load(f)
     except Exception as err:
-        logger.warning("Cannot open {} due to {}.".format(os.path.join(dataset_dir, DATA_CONFIG), str(err)))
+        logger.warning(f"Cannot open {os.path.join(dataset_dir, DATA_CONFIG)} due to {str(err)}.")
         return {}
 
 

src/llamafactory/webui/components/data.py

@@ -41,7 +41,7 @@ def next_page(page_index: int, total_num: int) -> int:
 
 def can_preview(dataset_dir: str, dataset: list) -> "gr.Button":
     try:
-        with open(os.path.join(dataset_dir, DATA_CONFIG), "r", encoding="utf-8") as f:
+        with open(os.path.join(dataset_dir, DATA_CONFIG), encoding="utf-8") as f:
             dataset_info = json.load(f)
     except Exception:
         return gr.Button(interactive=False)
@@ -57,7 +57,7 @@ def can_preview(dataset_dir: str, dataset: list) -> "gr.Button":
 
 
 def _load_data_file(file_path: str) -> List[Any]:
-    with open(file_path, "r", encoding="utf-8") as f:
+    with open(file_path, encoding="utf-8") as f:
         if file_path.endswith(".json"):
             return json.load(f)
         elif file_path.endswith(".jsonl"):
@@ -67,7 +67,7 @@ def _load_data_file(file_path: str) -> List[Any]:
 
 
 def get_preview(dataset_dir: str, dataset: list, page_index: int) -> Tuple[int, list, "gr.Column"]:
-    with open(os.path.join(dataset_dir, DATA_CONFIG), "r", encoding="utf-8") as f:
+    with open(os.path.join(dataset_dir, DATA_CONFIG), encoding="utf-8") as f:
         dataset_info = json.load(f)
 
     data_path = os.path.join(dataset_dir, dataset_info[dataset[0]]["file_name"])

src/llamafactory/webui/engine.py

@@ -56,9 +56,9 @@ class Engine:
         if not self.pure_chat:
             current_time = get_time()
             init_dict["train.current_time"] = {"value": current_time}
-            init_dict["train.output_dir"] = {"value": "train_{}".format(current_time)}
-            init_dict["train.config_path"] = {"value": "{}.yaml".format(current_time)}
-            init_dict["eval.output_dir"] = {"value": "eval_{}".format(current_time)}
+            init_dict["train.output_dir"] = {"value": f"train_{current_time}"}
+            init_dict["train.config_path"] = {"value": f"{current_time}.yaml"}
+            init_dict["eval.output_dir"] = {"value": f"eval_{current_time}"}
             init_dict["infer.mm_box"] = {"visible": False}
 
             if user_config.get("last_model", None):

src/llamafactory/webui/manager.py

@@ -29,7 +29,7 @@ class Manager:
         Adds elements to manager.
         """
         for elem_name, elem in elem_dict.items():
-            elem_id = "{}.{}".format(tab_name, elem_name)
+            elem_id = f"{tab_name}.{elem_name}"
             self._id_to_elem[elem_id] = elem
             self._elem_to_id[elem] = elem_id
 

src/llamafactory/webui/runner.py

@@ -231,7 +231,7 @@ class Runner:
         if get("train.ds_stage") != "none":
             ds_stage = get("train.ds_stage")
             ds_offload = "offload_" if get("train.ds_offload") else ""
-            args["deepspeed"] = os.path.join(DEFAULT_CACHE_DIR, "ds_z{}_{}config.json".format(ds_stage, ds_offload))
+            args["deepspeed"] = os.path.join(DEFAULT_CACHE_DIR, f"ds_z{ds_stage}_{ds_offload}config.json")
 
         return args
 
@@ -313,7 +313,7 @@ class Runner:
             if args.get("deepspeed", None) is not None:
                 env["FORCE_TORCHRUN"] = "1"
 
-            self.trainer = Popen("llamafactory-cli train {}".format(save_cmd(args)), env=env, shell=True)
+            self.trainer = Popen(f"llamafactory-cli train {save_cmd(args)}", env=env, shell=True)
             yield from self.monitor()
 
     def _form_config_dict(self, data: Dict["Component", Any]) -> Dict[str, Any]:

src/llamafactory/webui/utils.py

@@ -111,14 +111,14 @@ def gen_cmd(args: Dict[str, Any]) -> str:
     """
     cmd_lines = ["llamafactory-cli train "]
     for k, v in clean_cmd(args).items():
-        cmd_lines.append("    --{} {} ".format(k, str(v)))
+        cmd_lines.append(f"    --{k} {str(v)} ")
 
     if os.name == "nt":
         cmd_text = "`\n".join(cmd_lines)
     else:
         cmd_text = "\\\n".join(cmd_lines)
 
-    cmd_text = "```bash\n{}\n```".format(cmd_text)
+    cmd_text = f"```bash\n{cmd_text}\n```"
     return cmd_text
 
 
@@ -139,9 +139,9 @@ def get_eval_results(path: os.PathLike) -> str:
     r"""
     Gets scores after evaluation.
     """
-    with open(path, "r", encoding="utf-8") as f:
+    with open(path, encoding="utf-8") as f:
         result = json.dumps(json.load(f), indent=4)
-    return "```json\n{}\n```\n".format(result)
+    return f"```json\n{result}\n```\n"
 
 
 def get_time() -> str:
@@ -161,13 +161,13 @@ def get_trainer_info(output_path: os.PathLike, do_train: bool) -> Tuple[str, "gr
 
     running_log_path = os.path.join(output_path, RUNNING_LOG)
     if os.path.isfile(running_log_path):
-        with open(running_log_path, "r", encoding="utf-8") as f:
+        with open(running_log_path, encoding="utf-8") as f:
             running_log = f.read()
 
     trainer_log_path = os.path.join(output_path, TRAINER_LOG)
     if os.path.isfile(trainer_log_path):
         trainer_log: List[Dict[str, Any]] = []
-        with open(trainer_log_path, "r", encoding="utf-8") as f:
+        with open(trainer_log_path, encoding="utf-8") as f:
             for line in f:
                 trainer_log.append(json.loads(line))
 
@@ -193,7 +193,7 @@ def load_args(config_path: str) -> Optional[Dict[str, Any]]:
     Loads saved arguments.
     """
     try:
-        with open(config_path, "r", encoding="utf-8") as f:
+        with open(config_path, encoding="utf-8") as f:
             return safe_load(f)
     except Exception:
         return None
@@ -211,7 +211,7 @@ def list_config_paths(current_time: str) -> "gr.Dropdown":
     r"""
     Lists all the saved configuration files.
     """
-    config_files = ["{}.yaml".format(current_time)]
+    config_files = [f"{current_time}.yaml"]
     if os.path.isdir(DEFAULT_CONFIG_DIR):
         for file_name in os.listdir(DEFAULT_CONFIG_DIR):
             if file_name.endswith(".yaml") and file_name not in config_files:
@@ -224,7 +224,7 @@ def list_output_dirs(model_name: Optional[str], finetuning_type: str, current_ti
     r"""
     Lists all the directories that can resume from.
     """
-    output_dirs = ["train_{}".format(current_time)]
+    output_dirs = [f"train_{current_time}"]
     if model_name:
         save_dir = get_save_dir(model_name, finetuning_type)
         if save_dir and os.path.isdir(save_dir):

tests/e2e/test_train.py

@@ -61,7 +61,7 @@ OS_NAME = os.environ.get("OS_NAME", "")
     ],
 )
 def test_run_exp(stage: str, dataset: str):
-    output_dir = "train_{}".format(stage)
+    output_dir = f"train_{stage}"
     run_exp({"stage": stage, "dataset": dataset, "output_dir": output_dir, **TRAIN_ARGS})
     assert os.path.exists(output_dir)