中文羊驼大模型二期 v2.0

[ymcui]

本次更新推出Chinese-LLaMA-2-13B和Chinese-Alpaca-2-13B模型，添加多种下游应用适配等。

🚀 推出Chinese-LLaMA-2-13B和Chinese-Alpaca-2-13B

• 与7B系列模型相同，Chinese-LLaMA-2-13B使用120G中文语料进行增量预训练，Chinese-Alpaca-2-13B进一步使用了500万指令数据进行精调
• 上述模型在多个客观指标上超过了一代Plus-13B甚至是Plus-33B的效果，查看系统效果
• 主观效果请通过模型竞技场浏览（模型刚加入排名暂不稳定）：http://llm-arena.ymcui.com/

🤖 添加多种应用适配

• 添加了基于llama.cpp的服务启动和API访问方法（Add server examples for llama.cpp #59）
• 添加text-generation-webui支持（add support for text-generation-webui #65）
• 添加LangChain使用教程（Add support for LangChain #79）
• 添加privateGPT适配，利用大模型和本地文档进行交互（Add privateGPT support #103）

相关使用文档请查阅📚 GitHub Wiki

💻 添加了CFG采样方法（#91）

• Classifier-Free Guidance (CFG) sampling 是一种提升模型输出与prompt内容的一致性的采样方法。
• 在transformers推理脚本中（inference_hf.py与gradio_demo.py）添加了对CFG采样方法的支持。
• 通过设置系统提示语（system_prompt）、反向提示语（negative prompt）以及调整CFG强度（guidance_scale），可以实现对模型输出的多样化控制。

其他更新

• 添加了OpenAI API流式输出支持（Streaming openai api support  #43）
• 添加了4比特推理支持（add support for 4bit inference #47）
• Gradio Demo中添加了system prompt修改框（Add system prompt input to Gradio demo #60）
• 修复了预训练之后生成无法停止的问题（Fixed llama pre-training bug #61）
• 修正了默认解码超参（Update default decoding hyperparameters #69）
• 添加了基于Chinese-LLaMA-2-7B的量化结果对比（Add quantization results on Chinese-LLaMA-2-7B #74）
• 优化了C-Eval预测脚本日志输出（Add --verbose argument to the C-Eval script #75）

For English release note, please refer to Discussion.

---

This discussion was created from the release 中文羊驼大模型二期 v2.0.

[ymcui]

Release Note for v2.0

Here's the translation of the provided markdown into English:

---

This update introduces the Chinese-LLaMA-2-13B and Chinese-Alpaca-2-13B models, adding compatibility for various downstream applications and more.

🚀 Launching Chinese-LLaMA-2-13B and Chinese-Alpaca-2-13B

• Similar to the 7B series models, Chinese-LLaMA-2-13B uses 120G of Chinese corpus for incremental pre-training, while Chinese-Alpaca-2-13B further fine-tunes with 5 million instruction data.
• The aforementioned models surpass the performance of the first-generation Plus-13B and even the Plus-33B in multiple objective metrics. Check the system performance.
• For subjective results, please browse through the model arena (the model has just joined the rankings and is not yet stable): http://llm-arena.ymcui.com/

🤖 Added Compatibility for Various Applications

• Added service startup and API access methods based on llama.cpp (Add server examples for llama.cpp #59).
• Added support for text-generation-webui (add support for text-generation-webui #65).
• Added a tutorial for using LangChain (Add support for LangChain #79).
• Added compatibility for privateGPT, which interacts with large models and local documents (Add privateGPT support #103).

For related documentation, please refer to 📚 GitHub Wiki.

💻 Introduced CFG Sampling Method (#91)

• Classifier-Free Guidance (CFG) sampling is a method to enhance the consistency of model outputs with the content of the prompt.
• Support for the CFG sampling method has been added in the transformers inference scripts (inference_hf.py and gradio_demo.py).
• By setting system prompts (system_prompt), negative prompts (negative prompt), and adjusting CFG strength (guidance_scale), diversified control of model outputs can be achieved.

Other Updates

• Added support for OpenAI API streaming output (Streaming openai api support  #43).
• Introduced 4-bit inference support (add support for 4bit inference #47).
• Added a system prompt modification box in the Gradio Demo (Add system prompt input to Gradio demo #60).
• Fixed an issue where generation couldn't stop after pre-training (Fixed llama pre-training bug #61).
• Corrected the default decoding hyperparameters (Update default decoding hyperparameters #69).
• Added quantification results comparison based on Chinese-LLaMA-2-7B (Add quantization results on Chinese-LLaMA-2-7B #74).
• Optimized the C-Eval prediction script log output (Add --verbose argument to the C-Eval script #75).