v0.3.3: ModelScope Integration, Reward Server

New features

• Support loading pre-trained models from ModelScope Hub by @tastelikefeet in #1700
• Support launching a reward model server in demo API via specifying --stage=rm in api_demo.py
• Support using a reward model server in PPO training via specifying --reward_model_type api
• Support adjusting the shard size of exported models via the export_size argument

New models

• Base models
  • DeepseekLLM-Base (7B/67B)
  • Qwen (1.8B/72B)
• Instruct/Chat models
  • DeepseekLLM-Chat (7B/67B)
  • Qwen-Chat (1.8B/72B)
  • Yi-34B-Chat

New datasets

• Supervised fine-tuning datasets
  • Nectar dataset by @mlinmg in #1689
• Preference datasets
  • Nectar dataset by @mlinmg in #1689

Bug fix

• Improve get_current_device by @billvsme in #1690
• Improve web UI preview by @Samge0 in #1695
• Fix #1543 #1597 #1657 #1658 #1659 #1668 #1682 #1696 #1699 #1703 #1707 #1710

v0.3.2: Patch release

New features

• Support training GPTQ quantized model #729 #1481 #1545
• Support resuming reward model training #1567

Bug fix

• Change default PPO parameters by @hannlp in #1553
• Fix ChatGLM2&3 templates #1453 #1480
• Fix #1548 by @Outsider565 in #1544
• Fix #1263 #1550 #1558

v0.3.0: Full-Parameter RLHF

New features

• Support full-parameter RLHF training (RM & PPO)
• Refactor llmtuner core in #1525 by @hiyouga
• Better LLaMA Board: full-parameter RLHF and demo mode

New models

• Base models
  • ChineseLLaMA-1.3B
  • LingoWhale-8B
• Instruct/Chat models
  • ChineseAlpaca-1.3B
  • Zephyr-7B-Alpha/Beta

Bug fix

• Fix bugs in partial-parameter (freeze) tuning
• Fix #224 #336 #931 #936 #1011 #1489 #1494 #1507 #1514

v0.2.2: Patch release

Bug fix

• Fix the OOM issue in PPO training by @mmbwf in #424
• Fix fine-tuning arguments by @yyq in #1454
• Refactor constants and evaluation by @hiyouga
• Fix #1452 #1466 #1478

v0.2.1: Variant Models, NEFTune Trick

New features

• Support NEFTune trick for supervised fine-tuning by @anvie in #1252
• Support loading dataset in the sharegpt format - read data/readme for details
• Support generating multiple responses in demo API via the n parameter
• Support caching the pre-processed dataset files via the cache_path argument
• Better LLaMA Board (pagination, controls, etc.)
• Support push_to_hub argument #1088

New models

• Base models
  • ChatGLM3-6B-Base
  • Yi (6B/34B)
  • Mistral-7B
  • BlueLM-7B-Base
  • Skywork-13B-Base
  • XVERSE-65B
  • Falcon-180B
  • Deepseek-Coder-Base (1.3B/6.7B/33B)
• Instruct/Chat models
  • ChatGLM3-6B
  • Mistral-7B-Instruct
  • BlueLM-7B-Chat
  • Zephyr-7B
  • OpenChat-3.5
  • Yayi (7B/13B)
  • Deepseek-Coder-Instruct (1.3B/6.7B/33B)

New datasets

• Pre-training datasets
  • RedPajama V2
  • Pile
• Supervised fine-tuning datasets
  • OpenPlatypus
  • ShareGPT Hyperfiltered
  • ShareGPT4
  • UltraChat 200k
  • AgentInstruct
  • LMSYS Chat 1M
  • Evol Instruct V2

Bug fix

• Fix full-parameter DPO training #1383 #1422 (inspired by @mengban )
• Fix tokenizer config by @lvzii in #1436
• Fix #1197 #1215 #1217 #1218 #1228 #1232 #1285 #1287 #1290 #1316 #1325 #1349 #1356 #1365 #1411 #1418 #1438 #1439 #1446

v0.2.0: Web UI Refactor, LongLoRA

New features

• Support LongLoRA for the LLaMA models
• Support training the Qwen-14B and InternLM-20B models
• Support training state recovery for the all-in-one Web UI
• Support Ascend NPU by @statelesshz in #975
• Integrate MMLU, C-Eval and CMMLU benchmarks

Modifications

• Rename repository to LLaMA Factory (former LLaMA Efficient Tuning)
• Use the cutoff_len argument instead of max_source_length and max_target_length #944
• Add a train_on_prompt option #1184

Bug fix

• Fix numeric error caused by the layer norm dtype in 84b7486 [1]
• Fix bugs in PPO Trainer by @mmbwf in #900
• Fix #424 #762 #814 #887 #913 #1000 #1026 #1032 #1064 #1068 #1074 #1086 #1097 #1176 #1177 #1190 #1191

[1] huggingface/transformers#25598 (comment)

v0.1.8: FlashAttention-2 and Baichuan2

New features

• Support FlashAttention-2 for LLaMA models. (RTX4090, A100, A800 or H100 GPU is required)
• Support training the Baichuan2 models
• Use right-padding to avoid overflow in fp16 training (also mentioned here)
• Align the computation method of the reward score with DeepSpeed-Chat (better generation)
• Support --lora_target all argument which automatically finds the applicable modules for LoRA training

Bug fix

• Use efficient EOS tokens to align with the Baichuan training ( baichuan-inc/Baichuan2#23 )
• Remove PeftTrainer to save model checkpoints in DeepSpeed training
• Fix bugs in web UI by @beat4ocean in #596 by @codemayq in #644 #651 #678 #741 by @kinghuin in #786
• Add dataset explanation by @panpan0000 in #629
• Fix a bug in the DPO data collator
• Fix a bug of the ChatGLM2 tokenizer in right-padding
• #608 #617 #649 #757 #761 #763 #809 #818

v0.1.7: Script Preview and RoPE Scaling

New features

• Preview training script in Web UI by @codemayq in #479 #511
• Support resuming from checkpoints by @niuba in #434 (transformers>=4.31.0 required)
• Two RoPE scaling methods: linear and NTK-aware scaling for LLaMA models (transformers>=4.31.0 required)
• Support training the ChatGLM2-6B model
• Support PPO training in bfloat16 data type #551

Bug fix

• Unusual output of quantized models #278 #391
• Runtime error in distributed DPO training #480
• Unexpected truncation in generation #532
• Dataset streaming error in pre-training #548 #549
• Tensor shape mismatch in PPO training using ChatGLM2 #527 #528
• #475 #476 #478 #481 #494 #551

v0.1.6: DPO Training and Qwen-7B

• Adapt DPO training from the TRL library
• Support fine-tuning the Qwen-7B, Qwen-7B-Chat, XVERSE-13B, and ChatGLM2-6B models
• Implement the "safe" ChatML template for Qwen-7B-Chat
• Better Web UI
• Pretty readme by @codemayq #382
• New features: #395 #451
• Fix InternLM-7B inference #312
• Fix bugs: #351 #354 #361 #376 #408 #417 #420 #423 #426

v0.1.5: Patch release

• Fix LLaMA-2 template #307
• Fix bug in preprocessing 968ce0d
• Fix #294 #296