Giter VIP home page Giter VIP logo

chatglm2-sft's Introduction

chatglm2-6b-sft

chatglm2-6b, chatglm-6b微调/LORA/推理

踩坑(截至20230705)

1. torch>=2.0, 否则微调会报很多错误(单纯推理可以用低版本);
2. tokenizer.encode输出为 [gMASK, sop, 真实文本token]
    64789 = {str} '[MASK]'
    64790 = {str} '[gMASK]'
    64791 = {str} '[sMASK]'
    64792 = {str} 'sop'
    64793 = {str} 'eop'
3. modeling_chatglm.py自带get_masks()的代码full_attention_mask -= padding_mask.unsqueeze(-1) - 1改为
                full_attention_mask = full_attention_mask.long() - padding_mask.unsqueeze(-1).long() - 1
4. 不支持gradient_checkpointing, 修复的话需要modeling_chatglm.py新增get_input_embeddings, set_input_embeddings;
5. modeling_chatglm.py中的ChatGLMForConditionalGeneration类forward函数中的
      if full_attention_mask is None:  前加入  batch_size, seq_length = input_ids.shape
6. get_mask(), 一直以来都对chatglm的mask/position有一些疑惑;
    def get_masks(seq, bos_token_id=sop):
        """  code from model_chatglm.py  """
        if seq.count(bos_token_id) == 2:
            context_length = seq[2:].index(bos_token_id) + 2
        else:
            context_length = seq.index(bos_token_id)
        attention_mask = torch.ones((1, len(seq), len(seq)))
        attention_mask.tril_()
        attention_mask[..., :context_length] = 1
        # attention_mask.unsqueeze_(1)
        attention_mask = (attention_mask < 0.5).bool()
        return attention_mask
7. 严格按照官方prompt构建输入输出"注意[Round 1]很重要, 不能删减"):
    输入"[Round 1]\n\n问:{}\n\n答:"
    输出"{}"
    输入id: [gMASK, sop, 输入tokens, gMASK, sop]
    输出id: [输出tokens, EOS]
8. prompttokenizer测试结论“[gMASK, sop]”很重要, 必须置于开头, 不能删减问题1+1=
   拼接:[gMASK, sop] + token_ids_0 #### 答案1: 1+1=2。
   拼接token_ids_0 + [gMASK, sop] #### 答案2.1: Hello! How can I assist you today?
   拼接token_ids_0 + [gMASK, sop] #### 答案2.2: 【问题】 【解答: 没有,tected,没有。 解答:protected,所以没有。
   拼接:[gMASK, sop] + token_ids_0 + [gMASK, sop] #### 答案3.1: the sum of 1 and 1 is 2.
   拼接:[gMASK, sop] + token_ids_0 + [gMASK, sop] #### 答案3.2: of course, 1+1=2
   拼接:[gMASK, sop] + token_ids_0 + [gMASK, sop] #### 答案3.3: 1+1=2。
   拼接:[gMASK, sop] + token_ids_0 + [BOS=1] #### 答案4.1: ""
   拼接第一轮剔除"[Round 1]",[gMASK, sop] + token_ids_0 #### 答案5: a />

环境配置

transformers==4.27.1
torch>=2.0
sentencepiece
cpm_kernels
mdtex2html
accelerate
protobuf
gradio

微调样例

地址: chatglm2_6b/ft_chatglm2

配置: chatglm2_6b/ft_chatglm2/config.py
训练: python train.py
推理: python predict.py
验证: python evaluation.py
接口: python post_api.py

实验日志

微调日志(alpaca_gpt4forall)

chatglm2_6b/loss_gpt4forall_alpaca.png

推理样例(LoRA, R=8)

chatglm2_6b/predict_sample_1.png

chatglm2_6b/predict_sample_2.png

参考/感谢

免责申明

本项目相关资源仅供学术研究之用,使用涉及第三方代码的部分时,请严格遵循相应的开源协议。模型生成的内容受模型计算、随机性和量化精度损失等因素影响,本项目不对其准确性作出保证。对于模型输出的任何内容,本项目不承担任何法律责任,亦不对因使用相关资源和输出结果而可能产生的任何损失承担责任。

chatglm2-sft's People

Contributors

moyongzhuo avatar yongzhuo avatar

Recommend Projects

  • React photo React

    A declarative, efficient, and flexible JavaScript library for building user interfaces.

  • Vue.js photo Vue.js

    🖖 Vue.js is a progressive, incrementally-adoptable JavaScript framework for building UI on the web.

  • Typescript photo Typescript

    TypeScript is a superset of JavaScript that compiles to clean JavaScript output.

  • TensorFlow photo TensorFlow

    An Open Source Machine Learning Framework for Everyone

  • Django photo Django

    The Web framework for perfectionists with deadlines.

  • D3 photo D3

    Bring data to life with SVG, Canvas and HTML. 📊📈🎉

Recommend Topics

  • javascript

    JavaScript (JS) is a lightweight interpreted programming language with first-class functions.

  • web

    Some thing interesting about web. New door for the world.

  • server

    A server is a program made to process requests and deliver data to clients.

  • Machine learning

    Machine learning is a way of modeling and interpreting data that allows a piece of software to respond intelligently.

  • Game

    Some thing interesting about game, make everyone happy.

Recommend Org

  • Facebook photo Facebook

    We are working to build community through open source technology. NB: members must have two-factor auth.

  • Microsoft photo Microsoft

    Open source projects and samples from Microsoft.

  • Google photo Google

    Google ❤️ Open Source for everyone.

  • D3 photo D3

    Data-Driven Documents codes.