README.md

metadata

license: llama3
datasets:
  - REILX/extracted_tagengo_gpt4
  - TigerResearch/sft_zh
  - alexl83/AlpacaDataCleaned
  - LooksJuicy/ruozhiba
  - silk-road/alpaca-data-gpt4-chinese
  - databricks/databricks-dolly-15k
  - microsoft/orca-math-word-problems-200k
  - Sao10K/Claude-3-Opus-Instruct-5K
language:
  - zh
  - en
tags:
  - text-generation-inference
  - llama
  - chat
  - sft
  - lora

数据集

使用以下8个数据集 对Llama-3-8B-Instruct进行微调。

基础模型：

• https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct

训练工具

https://github.com/hiyouga/LLaMA-Factory

测评方式：

使用opencompass(https://github.com/open-compass/OpenCompass/ )， 测试工具基于CEval和MMLU对微调之后的模型和原始模型进行测试。
测试模型分别为：

• Llama-3-8B
• Llama-3-8B-Instruct
• Llama-3-8B-Instruct-750Mb-lora, 使用8DataSets数据集对Llama-3-8B-Instruct模型进行sft方式lora微调

测试机器

8*A800

8DataSets数据集：

大约750Mb的微调数据集

• https://huggingface.co/datasets/REILX/extracted_tagengo_gpt4
• https://huggingface.co/datasets/TigerResearch/sft_zh
• https://huggingface.co/datasets/silk-road/alpaca-data-gpt4-chinese
• https://huggingface.co/datasets/LooksJuicy/ruozhiba
• https://huggingface.co/datasets/microsoft/orca-math-word-problems-200k
• https://huggingface.co/datasets/alexl83/AlpacaDataCleaned
• https://huggingface.co/datasets/Sao10K/Claude-3-Opus-Instruct-5K

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 5e-05
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 300
• num_epochs: 1.0