Supervised Finetune Training

This is the Introduction of our repository, for the project-1 of natural language process (CS3602).

Result

Pure sft full finetune

Lora fine tune 1.5B

To developers

For those who maintaining the project, please be aware that you should work under the directories that you are respond to.

File Structure

root

.
├── Dataset     # 数据集存放
├── README.md   # README
├── model   # 模型存放
├── nlp-project1    # 项目一
├── nlp-project2    # 项目二
└── report.pdf  # 报告

nlp-project1

.
├── assets  # 实验的统计数据表格和结果图片
│   ├── Comparison.png
│   ├── Dataset_results.xlsx
│   ├── Lora_results_on_1.5B.xlsx
│   ├── Parameter.png
│   ├── all_lora.png
│   ├── classical_lora.png
│   └── parameter.xlsx
├── auto_clean.py   # 训练过程中清楚checkpoint脚本
├── eval.py     # 简单的评估脚本
├── evaluation_results  # opencompass评测结果
│   ├── LoRA_finetuned_eval
│   │   └── 20241231_210406
│   │       └── summary
│   │           ├── summary_20241231_210406.csv
│   │           ├── summary_20241231_210406.md
│   │           └── summary_20241231_210406.txt
│   ├── base_1.5B_eval
│   │   └── 20241231_200238
│   │       └── summary
│   │           ├── summary_20241231_200238.csv
│   │           ├── summary_20241231_200238.md
│   │           └── summary_20241231_200238.txt
│   ├── base_eval
│   │   └── 20241223_180428
│   │       └── summary
│   │           ├── summary_20241223_180428.csv
│   │           ├── summary_20241223_180428.md
│   │           └── summary_20241223_180428.txt
│   ├── evals_masked_sft
│   │   └── 20241223_174211
│   │       └── summary
│   │           ├── summary_20241223_174211.csv
│   │           ├── summary_20241223_174211.md
│   │           └── summary_20241223_174211.txt
│   └── evals_unmasked_sft
│       └── 20241224_171102
│           └── summary
│               ├── summary_20241224_171102.csv
│               ├── summary_20241224_171102.md
│               └── summary_20241224_171102.txt
├── finetune_masked.ipynb   # 全序列输出SFT微调
├── finetune_unmasked.ipynb # output-onlySFT微调
└── peft.ipynb  # peft lora微调

nlp-project2

.
├── README.md #相关的使用方法和注意事项
├── main.py #聊天机器人主程序
├── model #聊天机器人实现
│   ├── __init__.py
│   ├── api_model.py
│   ├── base.py
│   ├── character_settings.yaml
│   └── local_model.py
├── model_test #聊天机器人测试程序和结果
│   ├── ans_character.txt
│   ├── ans_db.txt
│   ├── ans_harm.txt
│   ├── ans_info.txt
│   ├── ans_know.txt
│   ├── test_character.py
│   ├── test_db.py
│   └── test_model.py
├── requirements.txt #环境配置文件
└── server #一个简单的聊天机器人服务器
    ├── server.py
    └── web
        ├── Chat.js
        ├── WebClientPage.html
        └── style.css

Finetuned model link

For Qwen2.5-0.5B:

We will provide links for our output-loss only finetuned model and whole sequence model by the following links(named masked and unmasked):

download links by Baidu NetDisk, password 1234

download links by SJTU jBox, students and staff only, no dataset and base model

For Qwen2.5-1.5B:

We will provide links for our lora finetuned model by the following links:

download links by SJTU jBox, students and staff only, no base model

Important notifications

1. The finetune-evaluation code is finetune_masked.ipynb and finetuned_unmasked.ipynb, for the previous one is for output loss only finetune and the second one is for whole sequence loss calculation.
2. The final evaluation results was in evaluation_results. The final statistical tables and pictures was in assets.
3. After you downloaded the model, you can put the model at the following structure: