README.md

---
license: llama3
datasets:
- LooksJuicy/ruozhiba
language:
- zh
---

## 基于ruozhiba对Llama-3-8B-Instruct进行微调。</br>
### 模型：</br>
- https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct

### 数据集：
- https://huggingface.co/datasets/LooksJuicy/ruozhiba

### 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。</br>
测试模型分别为：
- Llama-3-8B
- Llama-3-8B-Instruct
- LLama3-Instruct-sft-ruozhiba,使用ruozhiba数据对Llama-3-8B-Instruct使用sft方式lora微调

### 结果
| 模型名称                 | CEVAL | MMLU |
|--------------------------|-------|------|
| LLama3                   | 49.91 | 66.62|
| LLama3-Instruct          | 50.55 | 67.15|
| LLama3-Instruct-sft-ruozhiba-3epoch | 50.87 | 67.51|

### Training hyperparameters
The following hyperparameters were used during training:
- learning_rate: 5e-05
- train_batch_size: 1
- eval_batch_size: 1
- seed: 42
- distributed_type: multi-GPU
- num_devices: 8
- gradient_accumulation_steps: 2
- total_train_batch_size: 16
- total_eval_batch_size: 8
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
- lr_scheduler_type: cosine
- lr_scheduler_warmup_steps: 20
- num_epochs: 3.0
- mixed_precision_training: Native AMP