From ed5c75bd649da6551e0ed6dc803fe29fd13929f0 Mon Sep 17 00:00:00 2001 From: khazic Date: Thu, 25 Jul 2024 09:03:21 +0800 Subject: [PATCH 1/4] Added the reference address for TRL PPO details. Former-commit-id: ceba96f9ed121bb75b8e802d9b758871a94046f1 --- README.md | 3 ++- README_zh.md | 2 +- 2 files changed, 3 insertions(+), 2 deletions(-) diff --git a/README.md b/README.md index 80ab347f..61f482fc 100644 --- a/README.md +++ b/README.md @@ -47,7 +47,8 @@ Choose your path: ## Features - **Various models**: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc. -- **Integrated methods**: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO, DPO, KTO, ORPO, etc. +- **Integrated methods**: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO (The details of TRL PPO can refer to [this blog](https://newfacade.github.io/notes-on-reinforcement-learning/17-ppo-trl.html).), DPO, KTO, ORPO, etc. + - **Scalable resources**: 16-bit full-tuning, freeze-tuning, LoRA and 2/3/4/5/6/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ. - **Advanced algorithms**: GaLore, BAdam, DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ, PiSSA and Agent tuning. - **Practical tricks**: FlashAttention-2, Unsloth, RoPE scaling, NEFTune and rsLoRA. diff --git a/README_zh.md b/README_zh.md index 962dcf43..17f424b2 100644 --- a/README_zh.md +++ b/README_zh.md @@ -47,7 +47,7 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272 ## 项目特色 - **多种模型**：LLaMA、LLaVA、Mistral、Mixtral-MoE、Qwen、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。 -- **集成方法**：（增量）预训练、（多模态）指令监督微调、奖励模型训练、PPO 训练、DPO 训练、KTO 训练、ORPO 训练等等。 +- **集成方法**：（增量）预训练、（多模态）指令监督微调、奖励模型训练、PPO（有关TRL PPO的详细信息，请参阅[此博客](https://newfacade.github.io/notes-on-reinforcement-learning/17-ppo-trl.html)）、DPO 训练、KTO 训练、ORPO 训练等等。 - **多种精度**：16 比特全参数微调、冻结微调、LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ 的 2/3/4/5/6/8 比特 QLoRA 微调。 - **先进算法**：GaLore、BAdam、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ、PiSSA 和 Agent 微调。 - **实用技巧**：FlashAttention-2、Unsloth、RoPE scaling、NEFTune 和 rsLoRA。 From f38decfbafa97e7a56be3e8acf7b549e02cff99f Mon Sep 17 00:00:00 2001 From: hoshi-hiyouga Date: Fri, 26 Jul 2024 11:29:09 +0800 Subject: [PATCH 2/4] Update README.md Former-commit-id: f97beca23a1c79df38769b8dd40c9b19d4e5ef5c --- README.md | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/README.md b/README.md index 61f482fc..8e41d832 100644 --- a/README.md +++ b/README.md @@ -47,7 +47,7 @@ Choose your path: ## Features - **Various models**: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc. -- **Integrated methods**: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO (The details of TRL PPO can refer to [this blog](https://newfacade.github.io/notes-on-reinforcement-learning/17-ppo-trl.html).), DPO, KTO, ORPO, etc. +- **Integrated methods**: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO, DPO, KTO, ORPO, etc. - **Scalable resources**: 16-bit full-tuning, freeze-tuning, LoRA and 2/3/4/5/6/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ. - **Advanced algorithms**: GaLore, BAdam, DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ, PiSSA and Agent tuning. @@ -201,6 +201,9 @@ You also can add a custom chat template to [template.py](src/llamafactory/data/t | ORPO Training | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: | | SimPO Training | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: | +> [!TIP] +> The implementation details of PPO can be found in [this blog](https://newfacade.github.io/notes-on-reinforcement-learning/17-ppo-trl.html). + ## Provided Datasets
Pre-training datasets From d4e84b9a11eede64d3bd29e7f58df2fabb067b00 Mon Sep 17 00:00:00 2001 From: hoshi-hiyouga Date: Fri, 26 Jul 2024 11:29:28 +0800 Subject: [PATCH 3/4] Update README.md Former-commit-id: 1186ad53d43dace9dec335331dbe246f1c5a729b --- README.md | 1 - 1 file changed, 1 deletion(-) diff --git a/README.md b/README.md index 8e41d832..14af3f46 100644 --- a/README.md +++ b/README.md @@ -48,7 +48,6 @@ Choose your path: - **Various models**: LLaMA, LLaVA, Mistral, Mixtral-MoE, Qwen, Yi, Gemma, Baichuan, ChatGLM, Phi, etc. - **Integrated methods**: (Continuous) pre-training, (multimodal) supervised fine-tuning, reward modeling, PPO, DPO, KTO, ORPO, etc. - - **Scalable resources**: 16-bit full-tuning, freeze-tuning, LoRA and 2/3/4/5/6/8-bit QLoRA via AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ. - **Advanced algorithms**: GaLore, BAdam, DoRA, LongLoRA, LLaMA Pro, Mixture-of-Depths, LoRA+, LoftQ, PiSSA and Agent tuning. - **Practical tricks**: FlashAttention-2, Unsloth, RoPE scaling, NEFTune and rsLoRA. From ca3dac9fb34f94703a4fe37b128caf6cfd2d7d5f Mon Sep 17 00:00:00 2001 From: hoshi-hiyouga Date: Fri, 26 Jul 2024 11:30:57 +0800 Subject: [PATCH 4/4] Update README_zh.md Former-commit-id: 77e7bfee7967319da6b5cc72e88d9f6cafe065b2 --- README_zh.md | 5 ++++- 1 file changed, 4 insertions(+), 1 deletion(-) diff --git a/README_zh.md b/README_zh.md index 17f424b2..578d2960 100644 --- a/README_zh.md +++ b/README_zh.md @@ -47,7 +47,7 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272 ## 项目特色 - **多种模型**：LLaMA、LLaVA、Mistral、Mixtral-MoE、Qwen、Yi、Gemma、Baichuan、ChatGLM、Phi 等等。 -- **集成方法**：（增量）预训练、（多模态）指令监督微调、奖励模型训练、PPO（有关TRL PPO的详细信息，请参阅[此博客](https://newfacade.github.io/notes-on-reinforcement-learning/17-ppo-trl.html)）、DPO 训练、KTO 训练、ORPO 训练等等。 +- **集成方法**：（增量）预训练、（多模态）指令监督微调、奖励模型训练、PPO 训练、DPO 训练、KTO 训练、ORPO 训练等等。 - **多种精度**：16 比特全参数微调、冻结微调、LoRA 微调和基于 AQLM/AWQ/GPTQ/LLM.int8/HQQ/EETQ 的 2/3/4/5/6/8 比特 QLoRA 微调。 - **先进算法**：GaLore、BAdam、DoRA、LongLoRA、LLaMA Pro、Mixture-of-Depths、LoRA+、LoftQ、PiSSA 和 Agent 微调。 - **实用技巧**：FlashAttention-2、Unsloth、RoPE scaling、NEFTune 和 rsLoRA。 @@ -200,6 +200,9 @@ https://github.com/user-attachments/assets/e6ce34b0-52d5-4f3e-a830-592106c4c272 | ORPO 训练 | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: | | SimPO 训练 | :white_check_mark: | :white_check_mark: | :white_check_mark: | :white_check_mark: | +> [!TIP] +> 有关 PPO 的实现细节，请参考[此博客](https://newfacade.github.io/notes-on-reinforcement-learning/17-ppo-trl.html)。 + ## 数据集
预训练数据集