--- language: - zh - en pipeline_tag: text-generation inference: false --- # Baichuan-13B-Instruction ![](./alpachino.png) ## 介绍 Baichuan-13B-Instruction 为 Baichuan-13B 系列模型进行指令微调后的版本，预训练模型可见 [Baichuan-13B-Base](https://huggingface.co/baichuan-inc/Baichuan-13B-Base)。 ## Demo 如下是一个使用 gradio 的模型 demo ```python import gradio as gr from transformers import AutoTokenizer, AutoModelForCausalLM tokenizer = AutoTokenizer.from_pretrained("AlpachinoNLP/Baichuan-13B-Instruction",trust_remote_code=True,use_fast=False) model = AutoModelForCausalLM.from_pretrained("AlpachinoNLP/Baichuan-13B-Instruction",trust_remote_code=True ).half() model.cuda() def generate(histories, max_new_tokens=2048, do_sample = True, top_p = 0.95, temperature = 0.35, repetition_penalty=1.1): prompt = "" for history in histories: history_with_identity = "\nHuman:" + history[0] + "\n\nAssistant:" + history[1] prompt += history_with_identity input_ids = tokenizer(prompt, return_tensors="pt").input_ids.to(model.device) outputs = model.generate( input_ids = input_ids, max_new_tokens=max_new_tokens, early_stopping=True, do_sample=do_sample, top_p=top_p, temperature=temperature, repetition_penalty=repetition_penalty, ) rets = tokenizer.batch_decode(outputs, skip_special_tokens=True) generate_text = rets[0].replace(prompt, "") return generate_text with gr.Blocks() as demo: chatbot = gr.Chatbot() msg = gr.Textbox() clear = gr.Button("clear") def user(user_message, history): return "", history + [[user_message, ""]] def bot(history): print(history) bot_message = generate(history) history[-1][1] = bot_message return history msg.submit(user, [msg, chatbot], [msg, chatbot], queue=False).then( bot, chatbot, chatbot ) clear.click(lambda: None, None, chatbot, queue=False) if __name__ == "__main__": demo.launch(server_name="0.0.0.0") ``` ## 量化部署 Baichuan-13B 支持 int8 和 int4 量化，用户只需在推理代码中简单修改两行即可实现。请注意，如果是为了节省显存而进行量化，应加载原始精度模型到 CPU 后再开始量化；避免在 `from_pretrained` 时添加 `device_map='auto'` 或者其它会导致把原始精度模型直接加载到 GPU 的行为的参数。 使用 int8 量化 (To use int8 quantization): ```python model = AutoModelForCausalLM.from_pretrained("AlpachinoNLP/Baichuan-13B-Instruction", torch_dtype=torch.float16, trust_remote_code=True) model = model.quantize(8).cuda() ``` 同样的，如需使用 int4 量化 (Similarly, to use int4 quantization): ```python model = AutoModelForCausalLM.from_pretrained("AlpachinoNLP/Baichuan-13B-Instruction", torch_dtype=torch.float16, trust_remote_code=True) model = model.quantize(4).cuda() ``` ## 模型详情 ### 模型结构 整体模型基于Baichuan-13B，为了获得更好的推理性能，Baichuan-13B 使用了 ALiBi 线性偏置技术，相对于 Rotary Embedding 计算量更小，对推理性能有显著提升；与标准的 LLaMA-13B 相比，生成 2000 个 tokens 的平均推理速度 (tokens/s)，实测提升 31.6%： | Model | tokens/s | | ------------ | -------- | | LLaMA-13B | 19.4 | | Baichuan-13B | 25.4 | 具体参数和见下表 | 模型名称 | 隐含层维度 | 层数 | 头数 | 词表大小 | 总参数量 | 训练数据（tokens） | 位置编码 | 最大长度 | | ------------ | ---------- | ---- | ---- | -------- | -------------- | ------------------ | ----------------------------------------- | -------- | | Baichuan-7B | 4,096 | 32 | 32 | 64,000 | 7,000,559,616 | 1.2万亿 | [RoPE](https://arxiv.org/abs/2104.09864) | 4,096 | | Baichuan-13B | 5,120 | 40 | 40 | 64,000 | 13,264,901,120 | 1.4万亿 | [ALiBi](https://arxiv.org/abs/2108.12409) | 4,096 | ## 训练详情 数据集主要由三部分组成： * 在 [sharegpt_zh](https://huggingface.co/datasets/QingyiSi/Alpaca-CoT/tree/main/ShareGPT) 数据集中筛选的出 13k 高质量数据。 * [lima](https://huggingface.co/datasets/GAIR/lima) * 按照任务类型挑选的 2.3k 高质量中文数据集，每个任务类型的数据量在 100 条左右。 硬件：8*A40 ## 测评结果 ## [CMMLU](https://github.com/haonan-li/CMMLU) | Model 5-shot | STEM | Humanities | Social Sciences | Others | China Specific | Average | | ---------------------------------------------------------- | :-------: | :--------: | :-------------: | :------: | :------------: | :------: | | Baichuan-7B | 34.4 | 47.5 | 47.6 | 46.6 | 44.3 | 44.0 | | Vicuna-13B | 31.8 | 36.2 | 37.6 | 39.5 | 34.3 | 36.3 | | Chinese-Alpaca-Plus-13B | 29.8 | 33.4 | 33.2 | 37.9 | 32.1 | 33.4 | | Chinese-LLaMA-Plus-13B | 28.1 | 33.1 | 35.4 | 35.1 | 33.5 | 33.0 | | Ziya-LLaMA-13B-Pretrain | 29.0 | 30.7 | 33.8 | 34.4 | 31.9 | 32.1 | | LLaMA-13B | 29.2 | 30.8 | 31.6 | 33.0 | 30.5 | 31.2 | | moss-moon-003-base (16B) | 27.2 | 30.4 | 28.8 | 32.6 | 28.7 | 29.6 | | Baichuan-13B-Base | 41.7 | 61.1 | 59.8 | 59.0 | 56.4 | 55.3 | | Baichuan-13B-Chat | 42.8 | **62.6** | **59.7** | **59.0** | **56.1** | **55.8** | | **Baichuan-13B-Instruction** | **44.50** | 61.16 | 59.07 | 58.34 | 55.55 | 55.61 | | Model zero-shot | STEM | Humanities | Social Sciences | Others | China Specific | Average | | ------------------------------------------------------------ | :-------: | :--------: | :-------------: | :-------: | :------------: | :-------: | | [ChatGLM2-6B](https://huggingface.co/THUDM/chatglm2-6b) | 41.28 | 52.85 | 53.37 | 52.24 | 50.58 | 49.95 | | [Baichuan-7B](https://github.com/baichuan-inc/baichuan-7B) | 32.79 | 44.43 | 46.78 | 44.79 | 43.11 | 42.33 | | [ChatGLM-6B](https://github.com/THUDM/GLM-130B) | 32.22 | 42.91 | 44.81 | 42.60 | 41.93 | 40.79 | | [BatGPT-15B](https://arxiv.org/abs/2307.00360) | 33.72 | 36.53 | 38.07 | 46.94 | 38.32 | 38.51 | | [Chinese-LLaMA-13B](https://github.com/ymcui/Chinese-LLaMA-Alpaca) | 26.76 | 26.57 | 27.42 | 28.33 | 26.73 | 27.34 | | [MOSS-SFT-16B](https://github.com/OpenLMLab/MOSS) | 25.68 | 26.35 | 27.21 | 27.92 | 26.70 | 26.88 | | [Chinese-GLM-10B](https://github.com/THUDM/GLM) | 25.57 | 25.01 | 26.33 | 25.94 | 25.81 | 25.80 | | [Baichuan-13B](https://github.com/baichuan-inc/Baichuan-13B) | 42.04 | 60.49 | 59.55 | 56.60 | 55.72 | 54.63 | | [Baichuan-13B-Chat](https://github.com/baichuan-inc/Baichuan-13B) | 37.32 | 56.24 | 54.79 | 54.07 | 52.23 | 50.48 | | **Baichuan-13B-Instruction** | **42.56** | **62.09** | **60.41** | **58.97** | **56.95** | **55.88** | > 说明：CMMLU 是一个综合性的中文评估基准，专门用于评估语言模型在中文语境下的知识和推理能力。我们直接使用其官方的[评测脚本](https://github.com/haonan-li/CMMLU)对模型进行评测。Model zero-shot 表格中 [Baichuan-13B-Chat](https://github.com/baichuan-inc/Baichuan-13B) 的得分来自我们直接运行 CMMLU 官方的评测脚本得到，其他模型的的得分来自于 [CMMLU](https://github.com/haonan-li/CMMLU/tree/master) 官方的评测结果，Model 5-shot 中其他模型的得分来自于[Baichuan-13B](https://github.com/baichuan-inc/Baichuan-13B) 官方的评测结果。