中文羊驼大模型二期 v4.0

[ymcui]

本次更新推出长上下文版模型Chinese-LLaMA-2-7B-64K和Chinese-Alpaca-2-7B-64K，可直接应用于各类长文本下游任务。推出人类偏好对齐版模型Chinese-Alpaca-2-1.3B-RLHF和Chinese-Alpaca-2-7B-RLHF，大幅提升正确价值观表达能力。

🚀 Chinese-LLaMA-2-7B-64K、Chinese-Alpaca-2-7B-64K

• 通过YaRN方法，在增量数据上进行训练，使模型获得长文本处理能力
• 在长文本评测LongBench（中文+代码任务）上，64K版模型相比标准版有显著的性能提升，甚至优于16K版本模型，效果参阅长上下文版模型评测
• Chinese-Alpaca-2-7B-64K已加入到⚔️ 模型竞技场，欢迎体验评测

👮 Chinese-Alpaca-2-1.3B-RLHF、Chinese-Alpaca-2-7B-RLHF

• 以Chinese-Alpaca-2系列模型为基础，利用偏好数据和PPO算法进行人类偏好对齐精调，显著提升模型传递正确价值观的能力
• 在人类偏好对齐评测方面，RLHF模型表现显著超越对应的SFT模型，小模型（1.3B）超越大模型（7B），具体参阅人类偏好对齐（RLHF）版本评测
• Chinese-Alpaca-2-1.3B-RLHF、Chinese-Alpaca-2-7B-RLHF已加入到⚔️ 模型竞技场，欢迎体验评测

温馨提示

• 64K版长上下文模型已支持🤗HuggingFace、llama.cpp、langchain等生态
• 使用长上下文版模型之前，务必参阅推理与部署表格，查看相关方法是否支持
• 建议：主要处理长文本任务时，使用长上下文版模型；其他场景建议使用标准版模型

其他更新

• 更新了对高版本依赖的支持，通过参数设置选择全量训练或LoRA训练 (Update the code to support the latest dependency #393)
• 添加可直接下载使用的GGUF模型权重 (Add GGUF models (llama.cpp compatible) #411)
• 修复了开源peft版本与requirements冲突的BUG (Update requirements #468)

For English release note, please refer to Discussion.

---

This discussion was created from the release 中文羊驼大模型二期 v4.0.

[ymcui]

Release Note v4.0

This update introduces the long-context models Chinese-LLaMA-2-7B-64K and Chinese-Alpaca-2-7B-64K, which can be directly applied to various long-text downstream tasks. We also introduce the human preference-aligned models Chinese-Alpaca-2-1.3B-RLHF and Chinese-Alpaca-2-7B-RLHF, significantly enhancing the capability to express correct values.

🚀 Chinese-LLaMA-2-7B-64K, Chinese-Alpaca-2-7B-64K

• Trained on incremental data using the YaRN method, enabling the model to process long texts.
• On the long-text evaluation LongBench (Chinese + code tasks), the 64K version models show significant performance improvement over the standard versions, even surpassing the 16K version models. For details, see Long Context Version Model Evaluation.
• Chinese-Alpaca-2-7B-64K has been added to the ⚔️ Model Arena, and you are welcome to experience and evaluate it.

👮 Chinese-Alpaca-2-1.3B-RLHF, Chinese-Alpaca-2-7B-RLHF

• Based on the Chinese-Alpaca-2 series models, significantly enhanced the ability to convey correct values through fine-tuning with human preference data and the PPO algorithm.
• In human preference alignment evaluations, the RLHF models significantly outperform the corresponding SFT models, with the smaller model (1.3B) surpassing the larger model (7B). For details, see Human Preference Alignment (RLHF) Version Evaluation.
• Chinese-Alpaca-2-1.3B-RLHF and Chinese-Alpaca-2-7B-RLHF have been added to the ⚔️ Model Arena, and you are welcome to experience and evaluate them.

Friendly Reminder

• The 64K long-context models are now supported by 🤗HuggingFace, llama.cpp, langchain, and other ecosystems.
• Before using the long-context models, be sure to consult the Inference and Deployment table to check if the related methods are supported.
• Suggestion: Use the long-context models primarily for long-text tasks; for other scenarios, the standard models are recommended.

Other Updates

• Updated support for higher version dependencies, with a choice between full training or LoRA training through parameter settings (Update the code to support the latest dependency #393).
• Added GGUF model weights available for direct download and use (Add GGUF models (llama.cpp compatible) #411).
• Fixed a BUG where the open-source peft version conflicted with requirements (Update requirements #468).