add base model qlora fintuning config file and optimize deduplicate.py (#128)
This commit is contained in:
xzw
GitHub
commit
a12a7ef107
11
README.md
11
README.md
@ -38,7 +38,6 @@
|
||||
<a href="https://github.com/SmartFlowAI/EmoLLM/issues">提出新特性</a>
|
||||
</div>
|
||||
|
||||
|
||||
<!-- 本篇README.md面向开发者 -->
|
||||
|
||||
**EmoLLM** 是一系列能够支持 **理解用户-支持用户-帮助用户** 心理健康辅导链路的心理健康大模型，由 `LLM`指令微调而来，欢迎大家star~⭐⭐。目前已经开源的 `LLM` 微调配置如下：
|
||||
@ -49,6 +48,7 @@
|
||||
| :-------------------: | :--------: |
|
||||
| InternLM2_7B_chat | QLORA |
|
||||
| InternLM2_7B_chat | 全量微调 |
|
||||
| InternLM2_7B_base | QLORA |
|
||||
| InternLM2_1_8B_chat | 全量微调 |
|
||||
| InternLM2_20B_chat | LORA |
|
||||
| Qwen_7b_chat | QLORA |
|
||||
@ -110,6 +110,7 @@
|
||||
</details>
|
||||
|
||||
### 🏆荣誉栏
|
||||
|
||||
- 项目荣获上海人工智能实验室举办的**2024浦源大模型系列挑战赛春季赛*****50强***
|
||||
|
||||
<p align="center">
|
||||
@ -151,9 +152,10 @@
|
||||
- [如何参与本项目](#如何参与本项目)
|
||||
- [作者（排名不分先后）](#作者排名不分先后)
|
||||
- [版权说明](#版权说明)
|
||||
- [引用](#引用)
|
||||
- [特别鸣谢](#特别鸣谢)
|
||||
- [Star History](#star-history)
|
||||
- [🌟Contributors](#-contributors)
|
||||
- [🌟 Contributors](#-contributors)
|
||||
- [交流群](#交流群)
|
||||
|
||||
###### 开发前的配置要求
|
||||
@ -234,7 +236,7 @@ git clone https://github.com/SmartFlowAI/EmoLLM.git
|
||||
| [ZeyuBa](https://github.com/ZeyuBa) | 自动化所在读硕士 | | |
|
||||
| [aiyinyuedejustin](https://github.com/aiyinyuedejustin) | 宾夕法尼亚大学在读硕士 | | |
|
||||
| [Nobody-ML](https://github.com/Nobody-ML) | 中国石油大学（华东）在读本科生 | | |
|
||||
| [chg0901](https://github.com/chg0901) | [MiniSora](https://github.com/mini-sora/minisora/) |MiniSora主要维护| 数据清洗、文档翻译 |
|
||||
| [chg0901](https://github.com/chg0901) | [MiniSora](https://github.com/mini-sora/minisora/) |[MiniSora](https://github.com/mini-sora/minisora/)主要维护者，管理员| LLM微调、数据清洗、文档翻译 |
|
||||
| [Mxoder](https://github.com/Mxoder) | 北京航空航天大学在读本科生 | | |
|
||||
| [Anooyman](https://github.com/Anooyman) | 南京理工大学硕士 | | |
|
||||
| [Vicky-3021](https://github.com/Vicky-3021) | 西安电子科技大学硕士（研0） | | |
|
||||
@ -248,8 +250,8 @@ git clone https://github.com/SmartFlowAI/EmoLLM.git
|
||||
|
||||
该项目签署了 MIT 授权许可，详情请参阅 [LICENSE](https://github.com/SmartFlowAI/EmoLLM/blob/main/LICENSE)
|
||||
|
||||
|
||||
### 引用
|
||||
|
||||
如果本项目对您的工作有所帮助，请使用以下格式引用：
|
||||
|
||||
```bibtex
|
||||
@ -300,7 +302,6 @@ git clone https://github.com/SmartFlowAI/EmoLLM.git
|
||||
[OpenXLab_App-url]: https://openxlab.org.cn/apps/detail/Farewell1/EmoLLMV2.0
|
||||
[OpenXLab_Model-url]: https://openxlab.org.cn/models/detail/ajupyter/EmoLLM_internlm2_7b_full
|
||||
|
||||
|
||||
## 交流群
|
||||
|
||||
- 如果失效，请移步Issue区
|
||||
|
||||
10
README_EN.md
10
README_EN.md
@ -42,7 +42,6 @@
|
||||
|
||||
<!-- 本篇README.md面向开发者 -->
|
||||
|
||||
|
||||
**EmoLLM** is a series of large language models designed to understand, support and help customers in mental health counseling. It is fine-tuned from the LLM instructions. We really appreciate it if you could give it a star~⭐⭐. The open-sourced configuration is as follows:
|
||||
|
||||
<div align="center">
|
||||
@ -51,6 +50,7 @@
|
||||
| :-------------------: | :------: |
|
||||
| InternLM2_7B_chat | QLORA |
|
||||
| InternLM2_7B_chat | full fine-tuning |
|
||||
| InternLM2_7B_base | QLORA |
|
||||
| InternLM2_1_8B_chat | full fine-tuning |
|
||||
| InternLM2_20B_chat | LORA |
|
||||
| Qwen_7b_chat | QLORA |
|
||||
@ -90,7 +90,6 @@ The Model aims to fully understand and promote the mental health of individuals,
|
||||
|
||||
- 【2024.2.18】 The full fine-tuned version based on Qwen1_5-0_5B-Chat has been [open-sourced](https://www.modelscope.cn/models/aJupyter/EmoLLM_Qwen1_5-0_5B-Chat_full_sft/summary). Friends with limited computational resources can now dive in and explore it.
|
||||
|
||||
|
||||
<details>
|
||||
<summary>View More</summary>
|
||||
|
||||
@ -173,8 +172,6 @@ git clone https://github.com/SmartFlowAI/EmoLLM.git
|
||||
- [Deployment Guide](#deployment-guide)
|
||||
- View More Details
|
||||
|
||||
|
||||
|
||||
### File Directory Explanation
|
||||
|
||||
```
|
||||
@ -203,8 +200,8 @@ For details, see the [fine-tuning guide](xtuner_config/README.md)
|
||||
- Demo deployment: see [deployment guide](./demo/README.md) for details.
|
||||
- Quantitative deployment based on [LMDeploy](https://github.com/InternLM/lmdeploy/): see [deploy](./deploy/lmdeploy.md)
|
||||
|
||||
|
||||
### RAG (Retrieval Augmented Generation) Pipeline
|
||||
|
||||
- See [RAG](./rag/)
|
||||
|
||||
<details>
|
||||
@ -251,7 +248,7 @@ This project uses Git for version control. You can see the currently available v
|
||||
| [ZeyuBa](https://github.com/ZeyuBa) | Institute of Automation, Master's student | | |
|
||||
| [aiyinyuedejustin](https://github.com/aiyinyuedejustin) | University of Pennsylvania, Master's student | | |
|
||||
| [Nobody-ML](https://github.com/Nobody-ML) | China University of Petroleum (East China), Undergraduate student | | |
|
||||
| [chg0901](https://github.com/chg0901) | [MiniSora](https://github.com/mini-sora/minisora) |Maintainer and Admin| Data Cleaning and Docs Translation |
|
||||
| [chg0901](https://github.com/chg0901) | [MiniSora](https://github.com/mini-sora/minisora) |Maintainer and Admin of [MiniSora](https://github.com/mini-sora/minisora) | LLM Fine-Tuning, Data Cleaning and Docs Translation |
|
||||
| [Mxoder](https://github.com/Mxoder) | Beihang University, Undergraduate student | | |
|
||||
| [Anooyman](https://github.com/Anooyman) | Nanjing University of Science and Technology, Master's student | | |
|
||||
| [Vicky-3021](https://github.com/Vicky-3021) | Xidian University, Master's student (Research Year 0) | | |
|
||||
@ -308,6 +305,7 @@ The project is licensed under the MIT License. Please refer to the details
|
||||
[OpenXLab_Model-url]: https://openxlab.org.cn/models/detail/ajupyter/EmoLLM_internlm2_7b_full
|
||||
|
||||
## Communication group
|
||||
|
||||
- If it fails, go to the Issue section.
|
||||
|
||||
<p align="center">
|
||||
|
||||
@ -5,6 +5,9 @@ from datasketch import MinHash
|
||||
from hashlib import md5
|
||||
from simhash import Simhash
|
||||
|
||||
import time
|
||||
import numpy as np
|
||||
|
||||
def extract_text_from_json(obj, content):
|
||||
# print(content)
|
||||
if isinstance(obj, dict):
|
||||
@ -29,7 +32,7 @@ def is_duplicate_absolutely(d1, d2):
|
||||
def hash_dict(dict_obj):
|
||||
content = extract_text_from_json(dict_obj,'')
|
||||
content = content.replace('\n', '').replace('\t', '').replace(' ', '')
|
||||
print(content)
|
||||
# print(content)
|
||||
# m = get_minhash(content)
|
||||
m = Simhash(content)
|
||||
return m
|
||||
@ -43,10 +46,19 @@ def get_simhash(dict_obj):
|
||||
return Simhash(dict_obj)
|
||||
|
||||
# 使用绝对匹配和MinHash对dict列表去重
|
||||
def deduplicate_json(data_list, threshold=0.8):
|
||||
def deduplicate_json(data_list, threshold=0.8, time_print=True):
|
||||
seen_hashes = []
|
||||
keep = []
|
||||
duplicate = []
|
||||
|
||||
# global start
|
||||
start = time.time()
|
||||
last_start_seen_hashes = start
|
||||
last_start_duplicate = start
|
||||
stop1 = 0
|
||||
stop2 = 0
|
||||
print_interval = 500
|
||||
|
||||
for item in data_list:
|
||||
if not item['conversation']:
|
||||
continue
|
||||
@ -60,15 +72,36 @@ def deduplicate_json(data_list, threshold=0.8):
|
||||
has_similar = False
|
||||
# for stored_min_hash, stored_text in seen_hashes:
|
||||
# if stored_min_hash.jaccard(min_hash) > threshold:
|
||||
|
||||
for stored_min_hash, stored_text in seen_hashes:
|
||||
if 1 - (stored_min_hash.distance(sim_hash)/64.0) > threshold:
|
||||
has_similar = True
|
||||
duplicate.append(item)
|
||||
|
||||
print_len_duplicate = len(duplicate)+1
|
||||
if print_len_duplicate%print_interval == 0:
|
||||
if time_print:
|
||||
stop1 = time.time()
|
||||
print(f'print_len_duplicate={print_len_duplicate} Time: ', np.round(stop1 - last_start_duplicate, 5), np.round(stop1 - start , 5))
|
||||
last_start_duplicate = stop1
|
||||
else:
|
||||
print(f'print_len_duplicate={print_len_duplicate}')
|
||||
|
||||
break
|
||||
if not has_similar:
|
||||
# seen_hashes.append((min_hash,item))
|
||||
|
||||
seen_hashes.append((sim_hash,item))
|
||||
keep.append(item)
|
||||
|
||||
|
||||
print_len_seen_hashes = len(seen_hashes)+1
|
||||
if print_len_seen_hashes%print_interval == 0:
|
||||
if time_print:
|
||||
stop2 = time.time()
|
||||
print(f'print_len_seen_hashes={print_len_seen_hashes} Time: ', str(np.round(stop2 - last_start_seen_hashes,5)), str(np.round(stop2 - start, 5)))
|
||||
last_start_seen_hashes = stop2
|
||||
else:
|
||||
print(f'print_len_seen_hashes={print_len_seen_hashes}')
|
||||
else:
|
||||
duplicate.append(item)
|
||||
|
||||
@ -77,7 +110,8 @@ def deduplicate_json(data_list, threshold=0.8):
|
||||
|
||||
if __name__ == '__main__':
|
||||
DUP_THRESH = 0.8
|
||||
data_ai = 'qwen'
|
||||
data_ai = 'FatherLikeBF'
|
||||
# root_dir = rf'./datasets/{data_ai}/'
|
||||
root_dir = rf'./{data_ai}/'
|
||||
dedup_output_dir = os.path.join(root_dir,'dedup')
|
||||
if not os.path.exists(dedup_output_dir):
|
||||
@ -94,8 +128,13 @@ if __name__ == '__main__':
|
||||
with open(file_path, 'r', encoding='utf-8') as f:
|
||||
data = json.load(f)
|
||||
dedup_data, duplicate = deduplicate_json(data, DUP_THRESH)
|
||||
|
||||
with open(os.path.join(root_dir, 'dedup','dedup_' + file), 'w', encoding='utf-8') as output_file:
|
||||
json.dump(dedup_data, output_file, ensure_ascii=False, indent=4)
|
||||
|
||||
with open(os.path.join(root_dir, 'dedup','dup_' + file), 'w', encoding='utf-8') as output_file:
|
||||
json.dump(duplicate, output_file, ensure_ascii=False, indent=4)
|
||||
|
||||
for item in dedup_data:
|
||||
logger.info(f'dedup_data: {item}')
|
||||
for item in duplicate:
|
||||
|
||||
@ -1,17 +1,23 @@
|
||||
import torch
|
||||
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||||
from openxlab.model import download
|
||||
from modelscope import snapshot_download
|
||||
|
||||
download(model_repo='jujimeizuo/EmoLLM_Model',
|
||||
output='model')
|
||||
# download model in openxlab
|
||||
model_name_or_path =download(model_repo='ajupyter/EmoLLM_internlm2_7b_full',
|
||||
output='EmoLLM_internlm2_7b_full')
|
||||
|
||||
model_name_or_path = "model"
|
||||
# download model in modelscope
|
||||
model_name_or_path = snapshot_download('chg0901/EmoLLM-InternLM7B-base')
|
||||
|
||||
# offline model
|
||||
# model_name_or_path = "/root/StableCascade/emollm2/EmoLLM/xtuner_config/merged"
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_name_or_path, trust_remote_code=True)
|
||||
model = AutoModelForCausalLM.from_pretrained(model_name_or_path, trust_remote_code=True, torch_dtype=torch.bfloat16, device_map='auto')
|
||||
model = model.eval()
|
||||
|
||||
system_prompt = "你是一个由aJupyter、Farewell、jujimeizuo、Smiling&Weeping研发（排名按字母顺序排序，不分先后）、散步提供技术支持、上海人工智能实验室提供支持开发的心理健康大模型。现在你是一个心理专家，我有一些心理问题，请你用专业的知识帮我解决。"
|
||||
system_prompt = '你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。'
|
||||
|
||||
messages = [(system_prompt, '')]
|
||||
|
||||
|
||||
192
xtuner_config/README_internlm2_7b_base_qlora.md
Normal file
192
xtuner_config/README_internlm2_7b_base_qlora.md
Normal file
@ -0,0 +1,192 @@
|
||||
# InternLM2 7B Base QLoRA 微调指南
|
||||
|
||||
## 模型基座与配置文件
|
||||
|
||||
- 本项目在[**internlm2_7b_chat_qlora_e3**模型](./internlm2_7b_chat_qlora_e3.py)微调[指南](./README.md)的基础上，更新了对[**internlm2_7b_base_qlora_e3（配置文件）**](./internlm2_7b_base_qlora_e10_M_1e4_32_64.py)**模型**的微调。
|
||||
|
||||
## 数据
|
||||
|
||||
### 数据集
|
||||
|
||||
只采用了通用的数据集，不包括带有Role-Play色彩的数据集，详情请查看[数据集](../datasets/README.md)页面
|
||||
|
||||
| Category | Dataset | Type | Total |
|
||||
| :---------: | :-------------------: | :----------: | :-----: |
|
||||
| General | data | Conversation | 5600+ |
|
||||
| General | data_pro | Conversation | 36500+ |
|
||||
| General | multi_turn_dataset_1 | Conversation | 36,000+ |
|
||||
| General | multi_turn_dataset_2 | Conversation | 27,000+ |
|
||||
| General | single_turn_dataset_1 | QA | 14000+ |
|
||||
| General | single_turn_dataset_2 | QA | 18300+ |
|
||||
|
||||
### 数据集处理
|
||||
|
||||
#### 数据格式
|
||||
|
||||
训练用的数据，被处理成单轮或者多轮对话的形式，如
|
||||
|
||||
```python
|
||||
[
|
||||
# 多轮对话
|
||||
{
|
||||
"conversation": [
|
||||
{
|
||||
"input": "我最近很迷茫，高三后的压力让我感到恐惧，我对自己的存在感到怀疑，您能不能给我一些建议呢？",
|
||||
"output": "看到你面临的困境与压力，真心希望能够帮助你。首先，我想说的是，你并不是孤独的，高考前的困惑与迷茫是很常见的。不要给自己太大的压力，给自己短暂的放松。你认识到问题所在，这已经是一个很大的进步了。",
|
||||
"system": "你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。"
|
||||
},
|
||||
{
|
||||
"input": "谢谢您的鼓励，但我觉得自己学习成绩不好，没有进步，这让我更加困惑。",
|
||||
"output": "学习成绩好坏不是解决问题的关键，专注于自己的进步更为重要。为了达到自己的目标，制定一份合理的计划，按部就班的完成每一项任务。有时候自学不如求助他人，不要害怕问老师和同学问题，以便获得更好的帮助和理解。"
|
||||
},
|
||||
{
|
||||
"input": "我认为我缺乏自信，不知道怎样扭转自己的消极情绪，您能给我一些建议吗？",
|
||||
"output": "建立自信需要时间和努力。让我们从一点一滴的进步中建立自信心。设立一些具体的目标，完成这些目标将会培养自己的自信。鼓励自己，在每天结束前，总结一下自己的成就，这将帮助你更好的看到自身的优点。"
|
||||
},
|
||||
{
|
||||
"input": "我感到很无助，不知道该如何调整自己的心态。",
|
||||
"output": "了解自己的情绪并寻找有效的调节方法是处理感到无助情况的关键。比如爱好运动、听音乐和阅读书籍等，这些活动对于您的心理健康都是有积极帮助的。此外，在处理感到无助的情况下，家庭和朋友的支持也是非常关键的。"
|
||||
},
|
||||
{
|
||||
"input": "我感觉我的思维和学习方法都很糟糕，该怎样改进呢？",
|
||||
"output": "学习方法确实很重要，它会对你的成绩和情绪有很大的影响。最有效的学习方法是利用时间来做练习，多做一些题目，练习自己的掌握能力。建立起对知识的兴趣，这需要不断的积累，不要害怕失败或挫折，它们会成为助你成长的宝贵财富。"
|
||||
},
|
||||
{
|
||||
"input": "我会按照您的建议努力，但想问一下，我该如何面对可能的失败？",
|
||||
"output": "每个人都会面对失败和挫折，不要让它们打败你。碰到失败时，重新审视自己，分析分析自己的不足，寻找方法去改进。记得勇敢地看向未来，不要让过去的失败阻碍了自己前进的脚步。"
|
||||
},
|
||||
{
|
||||
"input": "非常感谢您的建议和支持，我会努力去实践。",
|
||||
"output": "这是我应该做的，你有勇气面对自己困难，你并不孤独，我们会一起为你打气，祝愿你早日获得心灵的平静和所有的成功。"
|
||||
}
|
||||
]
|
||||
},
|
||||
# 单轮对话
|
||||
{
|
||||
"conversation": [
|
||||
{
|
||||
"system": "你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。",
|
||||
"input": "一想到事情就紧张……?\n我最近因为考试，觉得心情很不好，自己没有及格，而且，明明觉得自己比室友付出的多，努力的多，平时做题也是一样的，可是不知道为什么只有我不及格，感觉很苦恼。而且我只要一有什么事情就十分紧张，考试也是，心跳加速，手心出汗，而且大脑一片空白，有时甚至会头疼，高考以及现在的考试都是这样。而且我觉得一情绪激动的时候就会咳嗽，想哭哭不出来，感觉很难受，头也会疼。就是觉得心情不好，但是有不敢表现出来。我该怎么办呢？",
|
||||
"output": "为什么不敢表现出来呢？试试表现出来，会怎样？\n有时候我们越抗拒，就会越强化。当我们接受了，反而症状就消失了。\n那我们要怎样接受呢？\n这本书《心灵的七种兵器》推荐看看哦，内容精简，又有真实事例可以帮助理解。\n我们为什么改不掉这个所谓的坏习惯。\n"
|
||||
}
|
||||
]
|
||||
},
|
||||
```
|
||||
|
||||
### 数据处理
|
||||
|
||||
- 使用 `../datasets/process.py` 以处理 **multi_turn_dataset(1 和 2，QA数据转单轮对话)**， `data.json` 和 `data_pro.json` 文件（两个多轮对话），以添加或者调整 **`system` prompt**
|
||||
- 使用 `../datasets/processed/process_single_turn_conversation_construction.py` 处理 **single-turn dataset** (1 和 2)，修改 (`input` 和 `ouput`) ，并在每次 **conversation** 中添加 **`system` prompt**
|
||||
- 使用 `../datasets/processed/process_merge.py` 用于合并 `../datasets/processed/` 目录下**6个更新后的数据集**，生成一个合并后的数据集 `combined_data.json`用于最终训练
|
||||
|
||||
### 数据量与训练epochs设置
|
||||
|
||||
- 由于采用了更大的数据集，我们对模型进行了**10 epoch**的训练，读者可以根据训练过程中的输出和loss变化，进行训练的终止和模型的挑选，也可以采用更加专业的评估方法，来对模型评测。
|
||||
- 在我们公布的托管于OpenXlab微调后的 internlm2_7b_chat_qlora微调模型中，我们保留了两个版本，一个是[5 epoch模型](https://openxlab.org.cn/models/detail/chg0901/EmoLLM-InternLM7B-base/tree/main)，另一个是[10 epoch模型](https://openxlab.org.cn/models/detail/chg0901/EmoLLM-InternLM7B-base-10e/tree/main)版本（**ModelScope**模型：[5 epoch模型](https://www.modelscope.cn/models/chg0901/EmoLLM-InternLM7B-base/files)和[10 epoch模型](https://www.modelscope.cn/models/chg0901/EmoLLM-InternLM7B-base-10e/files)）。
|
||||
|
||||
## 基于XTuner的微调🎉🎉🎉🎉🎉
|
||||
|
||||
### 环境准备
|
||||
|
||||
```markdown
|
||||
datasets==2.16.1
|
||||
deepspeed==0.13.1
|
||||
einops==0.7.0
|
||||
flash_attn==2.5.0
|
||||
openxlab==0.0.34
|
||||
peft==0.7.1
|
||||
sentencepiece==0.1.99
|
||||
torch==2.1.2
|
||||
transformers==4.36.2
|
||||
mmengine==0.10.3
|
||||
xtuner==0.1.15
|
||||
flash_attn==2.5.0
|
||||
```
|
||||
|
||||
也可以一键安装
|
||||
|
||||
```bash
|
||||
cd xtuner_config/
|
||||
pip install -r requirements.txt
|
||||
```
|
||||
|
||||
温馨提示：flash_attn的安装可能需要在本地编译，大约需要一到两小时，可以去[flash-attention](https://github.com/Dao-AILab/flash-attention/releases)中，查找和自己机器配置匹配的whl安装包或者采用InternLM AI studio提供的2.4.2版本whl安装包，自行安装，如：
|
||||
|
||||
```bash
|
||||
# from flash-attention
|
||||
pip install flash_attn-2.5.0+cu122torch2.1cxx11abiTRUE-cp310-cp310-linux_x86_64.whl
|
||||
|
||||
# from InternLM AI studio share folder
|
||||
pip install /root/share/wheels/flash_attn-2.4.2+cu118torch2.0cxx11abiTRUE-cp310-cp310-linux_x86_64.whl
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 微调
|
||||
|
||||
```bash
|
||||
cd xtuner_config/
|
||||
xtuner train internlm2_7b_base_qlora_e10_M_1e4_32_64.py --deepspeed deepspeed_zero2
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 将得到的 PTH 模型转换为 HuggingFace 模型
|
||||
|
||||
**即：生成 Adapter 文件夹**
|
||||
|
||||
```bash
|
||||
cd xtuner_config/
|
||||
mkdir hf
|
||||
export MKL_SERVICE_FORCE_INTEL=1
|
||||
|
||||
xtuner convert pth_to_hf internlm2_7b_base_qlora_e10_M_1e4_32_64.py ./work_dirs/internlm2_7b_base_qlora_e10_M_1e4_32_64/epoch_5.pth ./hf
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 将 HuggingFace adapter 合并到大语言模型
|
||||
|
||||
```bash
|
||||
xtuner convert merge /root/share/model_repos/internlm2-base-7b ./hf ./merged --max-shard-size 2GB
|
||||
# xtuner convert merge \
|
||||
# ${NAME_OR_PATH_TO_LLM} \
|
||||
# ${NAME_OR_PATH_TO_ADAPTER} \
|
||||
# ${SAVE_PATH} \
|
||||
# --max-shard-size 2GB
|
||||
```
|
||||
|
||||
### 10 epoch 模型的处理
|
||||
|
||||
```bash
|
||||
|
||||
cd xtuner_config/
|
||||
mkdir hf10
|
||||
export MKL_SERVICE_FORCE_INTEL=1
|
||||
|
||||
xtuner convert pth_to_hf internlm2_7b_base_qlora_e10_M_1e4_32_64.py ./work_dirs/internlm2_7b_base_qlora_e10_M_1e4_32_64/epoch_10.pth ./hf
|
||||
|
||||
xtuner convert merge /root/share/model_repos/internlm2-base-7b ./hf10 ./merged10 --max-shard-size 2GB
|
||||
# xtuner convert merge \
|
||||
# ${NAME_OR_PATH_TO_LLM} \
|
||||
# ${NAME_OR_PATH_TO_ADAPTER} \
|
||||
# ${SAVE_PATH} \
|
||||
# --max-shard-size 2GB
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
### 测试
|
||||
|
||||
```bash
|
||||
cd demo/
|
||||
python cli_internlm2.py
|
||||
```
|
||||
|
||||
---
|
||||
|
||||
## 其他
|
||||
|
||||
欢迎大家给[xtuner](https://github.com/InternLM/xtuner)和[EmoLLM](https://github.com/aJupyter/EmoLLM)点点star~
|
||||
|
||||
🎉🎉🎉🎉🎉
|
||||
210
xtuner_config/internlm2_7b_base_qlora_e10_M_1e4_32_64.py
Normal file
210
xtuner_config/internlm2_7b_base_qlora_e10_M_1e4_32_64.py
Normal file
@ -0,0 +1,210 @@
|
||||
# Copyright (c) OpenMMLab. All rights reserved.
|
||||
import torch
|
||||
from datasets import load_dataset
|
||||
from mmengine.dataset import DefaultSampler
|
||||
from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
|
||||
LoggerHook, ParamSchedulerHook)
|
||||
from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
|
||||
from peft import LoraConfig
|
||||
from torch.optim import AdamW
|
||||
from transformers import (AutoModelForCausalLM, AutoTokenizer,
|
||||
BitsAndBytesConfig)
|
||||
|
||||
from xtuner.dataset import process_hf_dataset
|
||||
from xtuner.dataset.collate_fns import default_collate_fn
|
||||
from xtuner.dataset.map_fns import template_map_fn_factory
|
||||
from xtuner.engine import DatasetInfoHook, EvaluateChatHook
|
||||
from xtuner.model import SupervisedFinetune
|
||||
from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
|
||||
|
||||
#######################################################################
|
||||
# PART 1 Settings #
|
||||
#######################################################################
|
||||
# Model
|
||||
# pretrained_model_name_or_path = '/root/share/model_repos/internlm2-chat-7b'
|
||||
pretrained_model_name_or_path = '/root/share/model_repos/internlm2-base-7b'
|
||||
|
||||
# Data
|
||||
# data_path = 'merge.json'
|
||||
data_path ='/root/StableCascade/emollm2/EmoLLM/datasets/processed/combined_data.json'
|
||||
|
||||
# https://github.com/InternLM/xtuner/blob/main/xtuner/utils/templates.py#L24C25-L24C25
|
||||
prompt_template = PROMPT_TEMPLATE.internlm2_chat # there is No internlm2_base
|
||||
|
||||
max_length = 2048
|
||||
pack_to_max_length = True
|
||||
|
||||
# Scheduler & Optimizer
|
||||
|
||||
# batch_size = 8 # per_device
|
||||
# accumulative_counts = 2
|
||||
batch_size = 16 # per_device
|
||||
accumulative_counts = 1
|
||||
|
||||
dataloader_num_workers = 0
|
||||
max_epochs = 10
|
||||
optim_type = AdamW
|
||||
lr = 1e-4
|
||||
betas = (0.9, 0.999)
|
||||
weight_decay = 0
|
||||
max_norm = 1 # grad clip
|
||||
warmup_ratio = 0.03
|
||||
|
||||
# Evaluate the generation performance during the training
|
||||
evaluation_freq = 500
|
||||
# SYSTEM = "现在你是一个心理专家，我有一些心理问题，请你用专业的知识帮我解决。"
|
||||
SYSTEM = "你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。"
|
||||
evaluation_inputs = [
|
||||
'我最近总是感到很焦虑，尤其是在学业上。我有个特别崇拜的同学，他好像在各方面都比我优秀，我总觉得自己怎么努力也追不上他，这让我压力特别大。',
|
||||
'我知道应该理性看待，但就是忍不住会去比较。我甚至晚上会因为这个睡不着觉，总想着怎样才能像他那样出色。',
|
||||
# ['我最近总是感到很焦虑，尤其是在学业上。我有个特别崇拜的同学，他好像在各方面都比我优秀，我总觉得自己怎么努力也追不上他，这让我压力特别大。',
|
||||
# '我知道应该理性看待，但就是忍不住会去比较。我甚至晚上会因为这个睡不着觉，总想着怎样才能像他那样出色。']
|
||||
]
|
||||
|
||||
#######################################################################
|
||||
# PART 2 Model & Tokenizer #
|
||||
#######################################################################
|
||||
tokenizer = dict(
|
||||
type=AutoTokenizer.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True,
|
||||
padding_side='right')
|
||||
|
||||
model = dict(
|
||||
type=SupervisedFinetune,
|
||||
llm=dict(
|
||||
type=AutoModelForCausalLM.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True,
|
||||
torch_dtype=torch.float16,
|
||||
quantization_config=dict(
|
||||
type=BitsAndBytesConfig,
|
||||
load_in_4bit=True,
|
||||
load_in_8bit=False,
|
||||
llm_int8_threshold=6.0,
|
||||
llm_int8_has_fp16_weight=False,
|
||||
bnb_4bit_compute_dtype=torch.float16,
|
||||
bnb_4bit_use_double_quant=True,
|
||||
bnb_4bit_quant_type='nf4')),
|
||||
lora=dict(
|
||||
type=LoraConfig,
|
||||
# r=64,
|
||||
# lora_alpha=16,
|
||||
r=32,
|
||||
lora_alpha=64,
|
||||
# r=16,
|
||||
# lora_alpha=32,
|
||||
lora_dropout=0.1,
|
||||
bias='none',
|
||||
task_type='CAUSAL_LM'))
|
||||
|
||||
#######################################################################
|
||||
# PART 3 Dataset & Dataloader #
|
||||
#######################################################################
|
||||
alpaca_en = dict(
|
||||
type=process_hf_dataset,
|
||||
dataset=dict(type=load_dataset, path='json', data_files=dict(train=data_path)),
|
||||
tokenizer=tokenizer,
|
||||
max_length=max_length,
|
||||
dataset_map_fn=None,
|
||||
template_map_fn=dict(
|
||||
type=template_map_fn_factory, template=prompt_template),
|
||||
remove_unused_columns=True,
|
||||
shuffle_before_pack=True,
|
||||
pack_to_max_length=pack_to_max_length)
|
||||
|
||||
train_dataloader = dict(
|
||||
batch_size=batch_size,
|
||||
num_workers=dataloader_num_workers,
|
||||
dataset=alpaca_en,
|
||||
sampler=dict(type=DefaultSampler, shuffle=True),
|
||||
collate_fn=dict(type=default_collate_fn))
|
||||
|
||||
#######################################################################
|
||||
# PART 4 Scheduler & Optimizer #
|
||||
#######################################################################
|
||||
# optimizer
|
||||
optim_wrapper = dict(
|
||||
type=AmpOptimWrapper,
|
||||
optimizer=dict(
|
||||
type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
|
||||
clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
|
||||
accumulative_counts=accumulative_counts,
|
||||
loss_scale='dynamic',
|
||||
dtype='float16')
|
||||
|
||||
# learning policy
|
||||
# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501
|
||||
param_scheduler = [
|
||||
dict(
|
||||
type=LinearLR,
|
||||
start_factor=1e-5,
|
||||
by_epoch=True,
|
||||
begin=0,
|
||||
end=warmup_ratio * max_epochs,
|
||||
convert_to_iter_based=True),
|
||||
dict(
|
||||
type=CosineAnnealingLR,
|
||||
eta_min=0.0,
|
||||
by_epoch=True,
|
||||
begin=warmup_ratio * max_epochs,
|
||||
T_max=max_epochs,
|
||||
convert_to_iter_based=True)
|
||||
]
|
||||
|
||||
# train, val, test setting
|
||||
train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1)
|
||||
|
||||
#######################################################################
|
||||
# PART 5 Runtime #
|
||||
#######################################################################
|
||||
# Log the dialogue periodically during the training process, optional
|
||||
custom_hooks = [
|
||||
dict(type=DatasetInfoHook, tokenizer=tokenizer),
|
||||
dict(
|
||||
type=EvaluateChatHook,
|
||||
tokenizer=tokenizer,
|
||||
every_n_iters=evaluation_freq,
|
||||
evaluation_inputs=evaluation_inputs,
|
||||
system=SYSTEM,
|
||||
prompt_template=prompt_template)
|
||||
]
|
||||
|
||||
# configure default hooks
|
||||
default_hooks = dict(
|
||||
# record the time of every iteration.
|
||||
timer=dict(type=IterTimerHook),
|
||||
# print log every 100 iterations.
|
||||
logger=dict(type=LoggerHook, interval=10),
|
||||
# enable the parameter scheduler.
|
||||
param_scheduler=dict(type=ParamSchedulerHook),
|
||||
# save checkpoint per epoch.
|
||||
checkpoint=dict(type=CheckpointHook, interval=1),
|
||||
# set sampler seed in distributed evrionment.
|
||||
sampler_seed=dict(type=DistSamplerSeedHook),
|
||||
)
|
||||
|
||||
# configure environment
|
||||
env_cfg = dict(
|
||||
# whether to enable cudnn benchmark
|
||||
cudnn_benchmark=False,
|
||||
# set multi process parameters
|
||||
mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
|
||||
# set distributed parameters
|
||||
dist_cfg=dict(backend='nccl'),
|
||||
)
|
||||
|
||||
# set visualizer
|
||||
visualizer = None
|
||||
|
||||
# set log level
|
||||
log_level = 'INFO'
|
||||
|
||||
# load from which checkpoint
|
||||
load_from = None
|
||||
|
||||
# whether to resume training from the loaded checkpoint
|
||||
resume = False
|
||||
|
||||
# Defaults to use random seed and disable `deterministic`
|
||||
randomness = dict(seed=None, deterministic=False)
|
||||
205
xtuner_config/internlm2_7b_base_qlora_e10_b8_16_32.py
Normal file
205
xtuner_config/internlm2_7b_base_qlora_e10_b8_16_32.py
Normal file
@ -0,0 +1,205 @@
|
||||
# Copyright (c) OpenMMLab. All rights reserved.
|
||||
import torch
|
||||
from datasets import load_dataset
|
||||
from mmengine.dataset import DefaultSampler
|
||||
from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
|
||||
LoggerHook, ParamSchedulerHook)
|
||||
from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
|
||||
from peft import LoraConfig
|
||||
from torch.optim import AdamW
|
||||
from transformers import (AutoModelForCausalLM, AutoTokenizer,
|
||||
BitsAndBytesConfig)
|
||||
|
||||
from xtuner.dataset import process_hf_dataset
|
||||
from xtuner.dataset.collate_fns import default_collate_fn
|
||||
from xtuner.dataset.map_fns import template_map_fn_factory
|
||||
from xtuner.engine import DatasetInfoHook, EvaluateChatHook
|
||||
from xtuner.model import SupervisedFinetune
|
||||
from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
|
||||
|
||||
#######################################################################
|
||||
# PART 1 Settings #
|
||||
#######################################################################
|
||||
# Model
|
||||
# pretrained_model_name_or_path = '/root/share/model_repos/internlm2-chat-7b'
|
||||
pretrained_model_name_or_path = '/root/share/model_repos/internlm2-base-7b'
|
||||
|
||||
# Data
|
||||
# data_path = 'merge.json'
|
||||
data_path ='/root/StableCascade/emollm2/EmoLLM/datasets/processed/combined_data.json'
|
||||
|
||||
# https://github.com/InternLM/xtuner/blob/main/xtuner/utils/templates.py#L24C25-L24C25
|
||||
prompt_template = PROMPT_TEMPLATE.internlm2_chat # there is No internlm2_base
|
||||
|
||||
max_length = 2048
|
||||
pack_to_max_length = True
|
||||
|
||||
# Scheduler & Optimizer
|
||||
|
||||
# batch_size = 8 # per_device
|
||||
# accumulative_counts = 2
|
||||
batch_size = 8 # per_device
|
||||
accumulative_counts = 1
|
||||
|
||||
dataloader_num_workers = 0
|
||||
max_epochs = 10
|
||||
optim_type = AdamW
|
||||
lr = 2e-4
|
||||
betas = (0.9, 0.999)
|
||||
weight_decay = 0
|
||||
max_norm = 1 # grad clip
|
||||
warmup_ratio = 0.03
|
||||
|
||||
# Evaluate the generation performance during the training
|
||||
evaluation_freq = 500
|
||||
# SYSTEM = "现在你是一个心理专家，我有一些心理问题，请你用专业的知识帮我解决。"
|
||||
SYSTEM = "你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。"
|
||||
evaluation_inputs = [
|
||||
'我最近总是感到很焦虑，尤其是在学业上。我有个特别崇拜的同学，他好像在各方面都比我优秀，我总觉得自己怎么努力也追不上他，这让我压力特别大。', '我知道应该理性看待，但就是忍不住会去比较。我甚至晚上会因为这个睡不着觉，总想着怎样才能像他那样出色。'
|
||||
]
|
||||
|
||||
#######################################################################
|
||||
# PART 2 Model & Tokenizer #
|
||||
#######################################################################
|
||||
tokenizer = dict(
|
||||
type=AutoTokenizer.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True,
|
||||
padding_side='right')
|
||||
|
||||
model = dict(
|
||||
type=SupervisedFinetune,
|
||||
llm=dict(
|
||||
type=AutoModelForCausalLM.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True,
|
||||
torch_dtype=torch.float16,
|
||||
quantization_config=dict(
|
||||
type=BitsAndBytesConfig,
|
||||
load_in_4bit=True,
|
||||
load_in_8bit=False,
|
||||
llm_int8_threshold=6.0,
|
||||
llm_int8_has_fp16_weight=False,
|
||||
bnb_4bit_compute_dtype=torch.float16,
|
||||
bnb_4bit_use_double_quant=True,
|
||||
bnb_4bit_quant_type='nf4')),
|
||||
lora=dict(
|
||||
type=LoraConfig,
|
||||
# r=64,
|
||||
# lora_alpha=16,
|
||||
r=16,
|
||||
lora_alpha=32,
|
||||
lora_dropout=0.1,
|
||||
bias='none',
|
||||
task_type='CAUSAL_LM'))
|
||||
|
||||
#######################################################################
|
||||
# PART 3 Dataset & Dataloader #
|
||||
#######################################################################
|
||||
alpaca_en = dict(
|
||||
type=process_hf_dataset,
|
||||
dataset=dict(type=load_dataset, path='json', data_files=dict(train=data_path)),
|
||||
tokenizer=tokenizer,
|
||||
max_length=max_length,
|
||||
dataset_map_fn=None,
|
||||
template_map_fn=dict(
|
||||
type=template_map_fn_factory, template=prompt_template),
|
||||
remove_unused_columns=True,
|
||||
shuffle_before_pack=True,
|
||||
pack_to_max_length=pack_to_max_length)
|
||||
|
||||
train_dataloader = dict(
|
||||
batch_size=batch_size,
|
||||
num_workers=dataloader_num_workers,
|
||||
dataset=alpaca_en,
|
||||
sampler=dict(type=DefaultSampler, shuffle=True),
|
||||
collate_fn=dict(type=default_collate_fn))
|
||||
|
||||
#######################################################################
|
||||
# PART 4 Scheduler & Optimizer #
|
||||
#######################################################################
|
||||
# optimizer
|
||||
optim_wrapper = dict(
|
||||
type=AmpOptimWrapper,
|
||||
optimizer=dict(
|
||||
type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
|
||||
clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
|
||||
accumulative_counts=accumulative_counts,
|
||||
loss_scale='dynamic',
|
||||
dtype='float16')
|
||||
|
||||
# learning policy
|
||||
# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501
|
||||
param_scheduler = [
|
||||
dict(
|
||||
type=LinearLR,
|
||||
start_factor=1e-5,
|
||||
by_epoch=True,
|
||||
begin=0,
|
||||
end=warmup_ratio * max_epochs,
|
||||
convert_to_iter_based=True),
|
||||
dict(
|
||||
type=CosineAnnealingLR,
|
||||
eta_min=0.0,
|
||||
by_epoch=True,
|
||||
begin=warmup_ratio * max_epochs,
|
||||
T_max=max_epochs,
|
||||
convert_to_iter_based=True)
|
||||
]
|
||||
|
||||
# train, val, test setting
|
||||
train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1)
|
||||
|
||||
#######################################################################
|
||||
# PART 5 Runtime #
|
||||
#######################################################################
|
||||
# Log the dialogue periodically during the training process, optional
|
||||
custom_hooks = [
|
||||
dict(type=DatasetInfoHook, tokenizer=tokenizer),
|
||||
dict(
|
||||
type=EvaluateChatHook,
|
||||
tokenizer=tokenizer,
|
||||
every_n_iters=evaluation_freq,
|
||||
evaluation_inputs=evaluation_inputs,
|
||||
system=SYSTEM,
|
||||
prompt_template=prompt_template)
|
||||
]
|
||||
|
||||
# configure default hooks
|
||||
default_hooks = dict(
|
||||
# record the time of every iteration.
|
||||
timer=dict(type=IterTimerHook),
|
||||
# print log every 100 iterations.
|
||||
logger=dict(type=LoggerHook, interval=10),
|
||||
# enable the parameter scheduler.
|
||||
param_scheduler=dict(type=ParamSchedulerHook),
|
||||
# save checkpoint per epoch.
|
||||
checkpoint=dict(type=CheckpointHook, interval=1),
|
||||
# set sampler seed in distributed evrionment.
|
||||
sampler_seed=dict(type=DistSamplerSeedHook),
|
||||
)
|
||||
|
||||
# configure environment
|
||||
env_cfg = dict(
|
||||
# whether to enable cudnn benchmark
|
||||
cudnn_benchmark=False,
|
||||
# set multi process parameters
|
||||
mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
|
||||
# set distributed parameters
|
||||
dist_cfg=dict(backend='nccl'),
|
||||
)
|
||||
|
||||
# set visualizer
|
||||
visualizer = None
|
||||
|
||||
# set log level
|
||||
log_level = 'INFO'
|
||||
|
||||
# load from which checkpoint
|
||||
load_from = None
|
||||
|
||||
# whether to resume training from the loaded checkpoint
|
||||
resume = False
|
||||
|
||||
# Defaults to use random seed and disable `deterministic`
|
||||
randomness = dict(seed=None, deterministic=False)
|
||||
203
xtuner_config/internlm2_7b_base_qlora_e3_M_1e4_32_64.py
Normal file
203
xtuner_config/internlm2_7b_base_qlora_e3_M_1e4_32_64.py
Normal file
@ -0,0 +1,203 @@
|
||||
# Copyright (c) OpenMMLab. All rights reserved.
|
||||
import torch
|
||||
from datasets import load_dataset
|
||||
from mmengine.dataset import DefaultSampler
|
||||
from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
|
||||
LoggerHook, ParamSchedulerHook)
|
||||
from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR, LinearLR
|
||||
from peft import LoraConfig
|
||||
from torch.optim import AdamW
|
||||
from transformers import (AutoModelForCausalLM, AutoTokenizer,
|
||||
BitsAndBytesConfig)
|
||||
|
||||
from xtuner.dataset import process_hf_dataset
|
||||
from xtuner.dataset.collate_fns import default_collate_fn
|
||||
from xtuner.dataset.map_fns import template_map_fn_factory
|
||||
from xtuner.engine import DatasetInfoHook, EvaluateChatHook
|
||||
from xtuner.model import SupervisedFinetune
|
||||
from xtuner.utils import PROMPT_TEMPLATE, SYSTEM_TEMPLATE
|
||||
|
||||
#######################################################################
|
||||
# PART 1 Settings #
|
||||
#######################################################################
|
||||
# Model
|
||||
# pretrained_model_name_or_path = '/root/share/model_repos/internlm2-chat-7b'
|
||||
pretrained_model_name_or_path = '/root/share/model_repos/internlm2-base-7b'
|
||||
|
||||
# Data
|
||||
# data_path = 'merge.json'
|
||||
data_path ='/root/StableCascade/emollm2/EmoLLM/datasets/processed/combined_data.json'
|
||||
|
||||
# https://github.com/InternLM/xtuner/blob/main/xtuner/utils/templates.py#L24C25-L24C25
|
||||
prompt_template = PROMPT_TEMPLATE.internlm2_chat # there is No internlm2_base
|
||||
|
||||
max_length = 2048
|
||||
pack_to_max_length = True
|
||||
|
||||
# Scheduler & Optimizer
|
||||
|
||||
# batch_size = 8 # per_device
|
||||
# accumulative_counts = 2
|
||||
batch_size = 16 # per_device
|
||||
accumulative_counts = 1
|
||||
|
||||
dataloader_num_workers = 0
|
||||
max_epochs = 3
|
||||
optim_type = AdamW
|
||||
lr = 1e-4
|
||||
betas = (0.9, 0.999)
|
||||
weight_decay = 0
|
||||
max_norm = 1 # grad clip
|
||||
warmup_ratio = 0.03
|
||||
|
||||
# Evaluate the generation performance during the training
|
||||
evaluation_freq = 500
|
||||
# SYSTEM = "现在你是一个心理专家，我有一些心理问题，请你用专业的知识帮我解决。"
|
||||
SYSTEM = "你是心理健康助手EmoLLM，由EmoLLM团队打造。你旨在通过专业心理咨询，协助来访者完成心理诊断。请充分利用专业心理学知识与咨询技术，一步步帮助来访者解决心理问题。"
|
||||
evaluation_inputs = [
|
||||
'我最近总是感到很焦虑，尤其是在学业上。我有个特别崇拜的同学，他好像在各方面都比我优秀，我总觉得自己怎么努力也追不上他，这让我压力特别大。', '我知道应该理性看待，但就是忍不住会去比较。我甚至晚上会因为这个睡不着觉，总想着怎样才能像他那样出色。'
|
||||
]
|
||||
|
||||
#######################################################################
|
||||
# PART 2 Model & Tokenizer #
|
||||
#######################################################################
|
||||
tokenizer = dict(
|
||||
type=AutoTokenizer.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True,
|
||||
padding_side='right')
|
||||
|
||||
model = dict(
|
||||
type=SupervisedFinetune,
|
||||
llm=dict(
|
||||
type=AutoModelForCausalLM.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True,
|
||||
torch_dtype=torch.float16,
|
||||
quantization_config=dict(
|
||||
type=BitsAndBytesConfig,
|
||||
load_in_4bit=True,
|
||||
load_in_8bit=False,
|
||||
llm_int8_threshold=6.0,
|
||||
llm_int8_has_fp16_weight=False,
|
||||
bnb_4bit_compute_dtype=torch.float16,
|
||||
bnb_4bit_use_double_quant=True,
|
||||
bnb_4bit_quant_type='nf4')),
|
||||
lora=dict(
|
||||
type=LoraConfig,
|
||||
r=32,
|
||||
lora_alpha=64,
|
||||
lora_dropout=0.1,
|
||||
bias='none',
|
||||
task_type='CAUSAL_LM'))
|
||||
|
||||
#######################################################################
|
||||
# PART 3 Dataset & Dataloader #
|
||||
#######################################################################
|
||||
alpaca_en = dict(
|
||||
type=process_hf_dataset,
|
||||
dataset=dict(type=load_dataset, path='json', data_files=dict(train=data_path)),
|
||||
tokenizer=tokenizer,
|
||||
max_length=max_length,
|
||||
dataset_map_fn=None,
|
||||
template_map_fn=dict(
|
||||
type=template_map_fn_factory, template=prompt_template),
|
||||
remove_unused_columns=True,
|
||||
shuffle_before_pack=True,
|
||||
pack_to_max_length=pack_to_max_length)
|
||||
|
||||
train_dataloader = dict(
|
||||
batch_size=batch_size,
|
||||
num_workers=dataloader_num_workers,
|
||||
dataset=alpaca_en,
|
||||
sampler=dict(type=DefaultSampler, shuffle=True),
|
||||
collate_fn=dict(type=default_collate_fn))
|
||||
|
||||
#######################################################################
|
||||
# PART 4 Scheduler & Optimizer #
|
||||
#######################################################################
|
||||
# optimizer
|
||||
optim_wrapper = dict(
|
||||
type=AmpOptimWrapper,
|
||||
optimizer=dict(
|
||||
type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
|
||||
clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
|
||||
accumulative_counts=accumulative_counts,
|
||||
loss_scale='dynamic',
|
||||
dtype='float16')
|
||||
|
||||
# learning policy
|
||||
# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501
|
||||
param_scheduler = [
|
||||
dict(
|
||||
type=LinearLR,
|
||||
start_factor=1e-5,
|
||||
by_epoch=True,
|
||||
begin=0,
|
||||
end=warmup_ratio * max_epochs,
|
||||
convert_to_iter_based=True),
|
||||
dict(
|
||||
type=CosineAnnealingLR,
|
||||
eta_min=0.0,
|
||||
by_epoch=True,
|
||||
begin=warmup_ratio * max_epochs,
|
||||
T_max=max_epochs,
|
||||
convert_to_iter_based=True)
|
||||
]
|
||||
|
||||
# train, val, test setting
|
||||
train_cfg = dict(by_epoch=True, max_epochs=max_epochs, val_interval=1)
|
||||
|
||||
#######################################################################
|
||||
# PART 5 Runtime #
|
||||
#######################################################################
|
||||
# Log the dialogue periodically during the training process, optional
|
||||
custom_hooks = [
|
||||
dict(type=DatasetInfoHook, tokenizer=tokenizer),
|
||||
dict(
|
||||
type=EvaluateChatHook,
|
||||
tokenizer=tokenizer,
|
||||
every_n_iters=evaluation_freq,
|
||||
evaluation_inputs=evaluation_inputs,
|
||||
system=SYSTEM,
|
||||
prompt_template=prompt_template)
|
||||
]
|
||||
|
||||
# configure default hooks
|
||||
default_hooks = dict(
|
||||
# record the time of every iteration.
|
||||
timer=dict(type=IterTimerHook),
|
||||
# print log every 100 iterations.
|
||||
logger=dict(type=LoggerHook, interval=10),
|
||||
# enable the parameter scheduler.
|
||||
param_scheduler=dict(type=ParamSchedulerHook),
|
||||
# save checkpoint per epoch.
|
||||
checkpoint=dict(type=CheckpointHook, interval=1),
|
||||
# set sampler seed in distributed evrionment.
|
||||
sampler_seed=dict(type=DistSamplerSeedHook),
|
||||
)
|
||||
|
||||
# configure environment
|
||||
env_cfg = dict(
|
||||
# whether to enable cudnn benchmark
|
||||
cudnn_benchmark=False,
|
||||
# set multi process parameters
|
||||
mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
|
||||
# set distributed parameters
|
||||
dist_cfg=dict(backend='nccl'),
|
||||
)
|
||||
|
||||
# set visualizer
|
||||
visualizer = None
|
||||
|
||||
# set log level
|
||||
log_level = 'INFO'
|
||||
|
||||
# load from which checkpoint
|
||||
load_from = None
|
||||
|
||||
# whether to resume training from the loaded checkpoint
|
||||
resume = False
|
||||
|
||||
# Defaults to use random seed and disable `deterministic`
|
||||
randomness = dict(seed=None, deterministic=False)
|
||||
@ -0,0 +1,222 @@
|
||||
# Copyright (c) OpenMMLab. All rights reserved.
|
||||
"""Data format:
|
||||
[
|
||||
{
|
||||
"conversation": [
|
||||
{
|
||||
"system": "",
|
||||
"input": "xxx",
|
||||
"output": "xxx"
|
||||
},
|
||||
{
|
||||
"input": "xxx",
|
||||
"output": "xxx"
|
||||
}
|
||||
]
|
||||
},
|
||||
...
|
||||
]
|
||||
Please refer to https://github.com/InternLM/xtuner/blob/main/docs/en/user_guides/dataset_format.md for details.
|
||||
""" # noqa: E501
|
||||
from datasets import load_dataset
|
||||
from mmengine.hooks import (CheckpointHook, DistSamplerSeedHook, IterTimerHook,
|
||||
LoggerHook, ParamSchedulerHook)
|
||||
from mmengine.optim import AmpOptimWrapper, CosineAnnealingLR
|
||||
from torch.optim import AdamW
|
||||
from torch.utils.data import BatchSampler
|
||||
from transformers import AutoModelForCausalLM, AutoTokenizer
|
||||
|
||||
from xtuner.dataset import process_hf_dataset
|
||||
from xtuner.dataset.collate_fns import default_collate_fn
|
||||
from xtuner.dataset.map_fns import template_map_fn_factory
|
||||
from xtuner.dataset.samplers import InternRepoSampler
|
||||
from xtuner.engine import (DatasetInfoHook, EvaluateChatHook, ThroughputHook,
|
||||
VarlenAttnArgsToMessageHubHook)
|
||||
from xtuner.engine.runner import TrainLoop
|
||||
from xtuner.model import SupervisedFinetune
|
||||
from xtuner.utils import PROMPT_TEMPLATE
|
||||
|
||||
#######################################################################
|
||||
# PART 1 Settings #
|
||||
#######################################################################
|
||||
# Model
|
||||
pretrained_model_name_or_path = 'internlm/internlm2-chat-7b'
|
||||
use_varlen_attn = True
|
||||
|
||||
# Data
|
||||
data_files = ['/path/to/json/file.json']
|
||||
prompt_template = PROMPT_TEMPLATE.internlm2_chat
|
||||
max_length = 32768
|
||||
pack_to_max_length = True
|
||||
|
||||
# Scheduler & Optimizer
|
||||
# batch size per device, set to 1 if `use_varlen_attn` = True
|
||||
# To clarify, enlarging the batch size essentially enlarges the `max_length`.
|
||||
# For example, doubling the max length is tantamount to doubling the batch size
|
||||
batch_size = 1
|
||||
accumulative_counts = 1 # 1bs * 1acc * 64gpu = 64 batchsize
|
||||
dataloader_num_workers = 4
|
||||
max_epochs = 1
|
||||
optim_type = AdamW
|
||||
lr = 4e-5
|
||||
betas = (0.9, 0.95)
|
||||
weight_decay = 0.01
|
||||
max_norm = 1 # grad clip
|
||||
warm_up_ratio = 0.025
|
||||
|
||||
# Save
|
||||
save_steps = 500
|
||||
save_total_limit = 2 # Maximum checkpoints to keep (-1 means unlimited)
|
||||
|
||||
# Evaluate the generation performance during the training
|
||||
evaluation_freq = 500
|
||||
SYSTEM = ''
|
||||
evaluation_inputs = [
|
||||
'请给我介绍五个上海的景点', 'Please tell me five scenic spots in Shanghai'
|
||||
]
|
||||
|
||||
#######################################################################
|
||||
# PART 2 Model & Tokenizer #
|
||||
#######################################################################
|
||||
tokenizer = dict(
|
||||
type=AutoTokenizer.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True,
|
||||
padding_side='right')
|
||||
|
||||
model = dict(
|
||||
type=SupervisedFinetune,
|
||||
use_varlen_attn=use_varlen_attn,
|
||||
llm=dict(
|
||||
type=AutoModelForCausalLM.from_pretrained,
|
||||
pretrained_model_name_or_path=pretrained_model_name_or_path,
|
||||
trust_remote_code=True))
|
||||
|
||||
#######################################################################
|
||||
# PART 3 Dataset & Dataloader #
|
||||
#######################################################################
|
||||
train_dataset = dict(
|
||||
type=process_hf_dataset,
|
||||
use_varlen_attn=use_varlen_attn,
|
||||
dataset=dict(type=load_dataset, path='json', data_files=data_files),
|
||||
tokenizer=tokenizer,
|
||||
max_length=max_length,
|
||||
dataset_map_fn=None,
|
||||
template_map_fn=dict(
|
||||
type=template_map_fn_factory, template=prompt_template),
|
||||
remove_unused_columns=True,
|
||||
shuffle_before_pack=True,
|
||||
pack_to_max_length=pack_to_max_length)
|
||||
|
||||
train_dataloader = dict(
|
||||
batch_size=batch_size,
|
||||
num_workers=dataloader_num_workers,
|
||||
dataset=train_dataset,
|
||||
sampler=dict(type=InternRepoSampler, shuffle=True, seed=1024),
|
||||
batch_sampler=dict(
|
||||
type=BatchSampler, drop_last=True, batch_size=batch_size),
|
||||
collate_fn=dict(type=default_collate_fn, use_varlen_attn=use_varlen_attn))
|
||||
|
||||
#######################################################################
|
||||
# PART 4 Scheduler & Optimizer #
|
||||
#######################################################################
|
||||
# optimizer
|
||||
optim_wrapper = dict(
|
||||
type=AmpOptimWrapper,
|
||||
optimizer=dict(
|
||||
type=optim_type, lr=lr, betas=betas, weight_decay=weight_decay),
|
||||
clip_grad=dict(max_norm=max_norm, error_if_nonfinite=False),
|
||||
accumulative_counts=accumulative_counts,
|
||||
loss_scale='dynamic',
|
||||
)
|
||||
|
||||
# learning policy
|
||||
# More information: https://github.com/open-mmlab/mmengine/blob/main/docs/en/tutorials/param_scheduler.md # noqa: E501
|
||||
param_scheduler = [
|
||||
dict(
|
||||
type='LinearLR',
|
||||
start_factor=1 / 40,
|
||||
by_epoch=True,
|
||||
begin=0,
|
||||
end=warm_up_ratio * max_epochs,
|
||||
convert_to_iter_based=True),
|
||||
dict(
|
||||
type=CosineAnnealingLR,
|
||||
eta_min=lr * 0.15,
|
||||
by_epoch=True,
|
||||
begin=warm_up_ratio * max_epochs,
|
||||
end=max_epochs,
|
||||
convert_to_iter_based=True)
|
||||
]
|
||||
|
||||
# train, val, test setting
|
||||
train_cfg = dict(type=TrainLoop, max_epochs=max_epochs)
|
||||
|
||||
#######################################################################
|
||||
# PART 5 Runtime #
|
||||
#######################################################################
|
||||
# Log the dialogue periodically during the training process, optional
|
||||
custom_hooks = [
|
||||
dict(
|
||||
type=DatasetInfoHook, tokenizer=tokenizer,
|
||||
is_intern_repo_dataset=True),
|
||||
dict(
|
||||
type=EvaluateChatHook,
|
||||
tokenizer=tokenizer,
|
||||
every_n_iters=evaluation_freq,
|
||||
evaluation_inputs=evaluation_inputs,
|
||||
system=SYSTEM,
|
||||
prompt_template=prompt_template),
|
||||
dict(type=ThroughputHook)
|
||||
]
|
||||
|
||||
if use_varlen_attn:
|
||||
custom_hooks += [dict(type=VarlenAttnArgsToMessageHubHook)]
|
||||
|
||||
# configure default hooks
|
||||
default_hooks = dict(
|
||||
# record the time of every iteration.
|
||||
timer=dict(type=IterTimerHook),
|
||||
# print log every 100 iterations.
|
||||
logger=dict(type=LoggerHook, log_metric_by_epoch=False, interval=1),
|
||||
# enable the parameter scheduler.
|
||||
param_scheduler=dict(type=ParamSchedulerHook),
|
||||
# save checkpoint per `save_steps`.
|
||||
checkpoint=dict(
|
||||
type=CheckpointHook,
|
||||
by_epoch=False,
|
||||
interval=save_steps,
|
||||
max_keep_ckpts=save_total_limit),
|
||||
# set sampler seed in distributed evrionment.
|
||||
sampler_seed=dict(type=DistSamplerSeedHook),
|
||||
)
|
||||
|
||||
# configure environment
|
||||
env_cfg = dict(
|
||||
# whether to enable cudnn benchmark
|
||||
cudnn_benchmark=False,
|
||||
# set multi process parameters
|
||||
mp_cfg=dict(mp_start_method='fork', opencv_num_threads=0),
|
||||
# set distributed parameters
|
||||
dist_cfg=dict(backend='nccl'),
|
||||
)
|
||||
|
||||
# set visualizer
|
||||
visualizer = None
|
||||
|
||||
# set log level
|
||||
log_level = 'INFO'
|
||||
|
||||
# load from which checkpoint
|
||||
load_from = None
|
||||
|
||||
# whether to resume training from the loaded checkpoint
|
||||
resume = False
|
||||
|
||||
# Defaults to use random seed and disable `deterministic`
|
||||
randomness = dict(seed=None, deterministic=False)
|
||||
|
||||
log_processor = dict(
|
||||
by_epoch=False,
|
||||
window_size=1,
|
||||
mean_pattern=r'.*(loss|time|data_time|grad_norm|tflops).*')
|
||||
10
xtuner_config/upload_modelscope.py
Normal file
10
xtuner_config/upload_modelscope.py
Normal file
@ -0,0 +1,10 @@
|
||||
from modelscope.hub.api import HubApi
|
||||
|
||||
YOUR_ACCESS_TOKEN = '请从ModelScope个人中心->访问令牌获取'
|
||||
|
||||
api = HubApi()
|
||||
api.login(YOUR_ACCESS_TOKEN)
|
||||
api.push_model(
|
||||
model_id="yourname/your_model_id",
|
||||
model_dir="my_model_dir" # 本地模型目录，要求目录中必须包含configuration.json
|
||||
)
|
||||
Loading…
Reference in New Issue
Block a user