基础环境安装

ENV_NAME="finbot"
PYTHON_VERSION="3.10"

## 创建虚拟环境并激活
conda create -n $ENV_NAME python=$PYTHON_VERSION notebook -y
conda activate $ENV_NAME

## 设置pip国内源
export PIP_INDEX_URL=https://pypi.tuna.tsinghua.edu.cn/simple

## 安装基础依赖
pip install \
  addict simplejson sortedcontainers openpyxl matplotlib \
  vllm peft FlagEmbedding bitsandbytes \
  catboost xgboost polars_ta lightgbm \
  modelscope hf_transfer \

## 安装评测依赖
pip install segeval backtrader deepeval 

## 安装LLM微调框架依赖--LLaMAfactory
cd tools
git clone --depth 1 https://github.com/hiyouga/LLaMA-Factory.git
cd LLaMA-Factory
pip install -e ".[torch,metrics]"
pip install deepspeed==0.15.4 # 此处版本号重要! 高版本有坑
llamafactory-cli version
cd ..

## 安装rag框架依赖--minirag, 该库未来应该可以通过pip install minirag-hku直接安装
cd tools
git clone https://github.com/HKUDS/MiniRAG.git
cd MiniRAG && pip install -e . && cd .. 
cd ..

部署LLM服务

通过vllm server来启动大模型，在终端输入：

CUDA_VISIBLE_DEVICES=0 \
vllm serve resources/open_models/Qwen2.5-3B-Instruct --trust-remote-code \
--served-model-name base   \
--enable-lora --lora-modules lora=resources/ckpts/Qwen2.5-3B-Instruct/lora_adapter \
--max-model-len 5000 --max-num-seqs 16 --quantization fp8 --gpu-memory-utilization 0.25 \
--port 12239

CUDA_VISIBLE_DEVICES=0,1 \
vllm serve resources/open_models/Qwen2.5-14B-Instruct --trust-remote-code \
--served-model-name judger  \
--max-model-len 5000 --max-num-seqs 30 --quantization fp8 --kv-cache-dtype fp8 \
--gpu-memory-utilization 0.4  --tensor-parallel-size 2 \
--port 12235 

--tensor-parallel-size 2 --pipeline-parallel-size 2

模型基座

• modelscope
  • AI-ModelScope/bge-large-zh-v1.5
  • Qwen/Qwen2.5-3B-Instruct
  • qolaris/FinBert
  • iic/nlp_bert_document-segmentation_chinese-base
  • Qwen/Qwen2.5-14B-Instruct
  • LLM-Research/Llama-3.2-3B-Instruct

终端下载示例：

MODEL=Qwen/Qwen2.5-14B-Instruct
LOCAL_PATH=resources/open_models/Qwen2.5-14B-Instruct
# 注意本地路径最后一级会直接作为模型文件夹
modelscope download --model $MODEL --local_dir $LOCAL_PATH

• huggingface
  • openbmb/MiniCPM3-4B
  • ProsusAI/finbert

终端下载示例：

MODEL=ProsusAI/finbert
LOCAL_PATH=resources/open_models/ProsusAI-finbert
# 注意本地路径最后一级会直接作为模型文件夹
huggingface-cli download --resume-download --local-dir-use-symlinks False $MODEL --local-dir $LOCAL_PATH

数据源

• 金融情绪提取
• 股价预测--市场数据
• 股价预测--新闻数据(自爬)
• 新闻摘要生成
• 文档主题分割

LLM微调

这一部分主要根据qwen2.5训练文档改写而来。

1. 数据准备

按照alpaca或sharegpt格式之一准备和注册数据集:

• alpaca格式:

  • 单条数据形式

    [
      {
        "instruction": "user instruction (required)",
        "input": "user input (optional)",
        "output": "model response (required)",
        "system": "system prompt (optional)",
        "history": [
          ["user instruction in the first round (optional)", "model response in the first round (optional)"],
          ["user instruction in the second round (optional)", "model response in the second round (optional)"]
        ]
      }
    ]

  • 在$resources/dataset_info.json中注册

    "dataset_name": {
      "file_name": "path/to/dataset",
      "columns": {
        "prompt": "instruction",
        "query": "input",
        "response": "output",
        "system": "system",
        "history": "history"
      }
    }

• sharegpt:

  • 单条数据形式

    [
      {
        "messages": [
          {"from": "user", "value": "user instruction"},
          {"from": "assistant", "value": "model response"}
        ],
        "system": "system prompt (optional)",
        "tools": "tool description (optional)"
      }
    ]

  • 在$resources/dataset_info.json中添加

    "dataset_name": {
          "file_name": "path/to/dataset",
          "formatting": "sharegpt",
          "columns": {
            "messages": "messages"
          },
          "tags": {
            "role_tag": "role",
            "content_tag": "content",
            "user_tag": "user",
            "assistant_tag": "assistant",
            "system_tag": "system"
          }
      }

2. 启动训练

训练脚本存放为$project-root/dev/sft_*.sh的形式

调整好其中的参数后从本项目根目录通过以下命令启动:

source train/sft_qwen2_5_3B_for_FINNA.sh

Note: 注意DATA_SETTINGS中template参数与所选模型匹配,详见https://github.com/hiyouga/LLaMA-Factory?tab=readme-ov-file#supported-models