Update README.md
Browse files
README.md
CHANGED
|
@@ -1,3 +1,42 @@
|
|
| 1 |
-
---
|
| 2 |
-
license: llama3
|
| 3 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: llama3
|
| 3 |
+
datasets:
|
| 4 |
+
- silk-road/alpaca-data-gpt4-chinese
|
| 5 |
+
- TigerResearch/sft_zh
|
| 6 |
+
- LooksJuicy/ruozhiba
|
| 7 |
+
- leo009/alpaca-cleaned-zh-cn
|
| 8 |
+
- REILX/extracted_tagengo_gpt4
|
| 9 |
+
language:
|
| 10 |
+
- en
|
| 11 |
+
- zh
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
### 模型：
|
| 15 |
+
- https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct
|
| 16 |
+
|
| 17 |
+
### 数据集：
|
| 18 |
+
- https://huggingface.co/datasets/TigerResearch/sft_zh
|
| 19 |
+
- https://huggingface.co/datasets/silk-road/alpaca-data-gpt4-chinese
|
| 20 |
+
- https://huggingface.co/datasets/REILX/extracted_tagengo_gpt4
|
| 21 |
+
- https://huggingface.co/datasets/LooksJuicy/ruozhiba
|
| 22 |
+
- https://huggingface.co/datasets/leo009/alpaca-cleaned-zh-cn
|
| 23 |
+
（使用langid清理以上数据集，删除其中非中文资料）
|
| 24 |
+
|
| 25 |
+
### 训练工具
|
| 26 |
+
https://github.com/hiyouga/LLaMA-Factory
|
| 27 |
+
|
| 28 |
+
### Training hyperparameters
|
| 29 |
+
The following hyperparameters were used during training:
|
| 30 |
+
- learning_rate: 5e-05
|
| 31 |
+
- train_batch_size: 4
|
| 32 |
+
- eval_batch_size: 8
|
| 33 |
+
- seed: 42
|
| 34 |
+
- distributed_type: multi-GPU
|
| 35 |
+
- num_devices: 8
|
| 36 |
+
- gradient_accumulation_steps: 4
|
| 37 |
+
- total_train_batch_size: 128
|
| 38 |
+
- total_eval_batch_size: 64
|
| 39 |
+
- optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
|
| 40 |
+
- lr_scheduler_type: cosine
|
| 41 |
+
- lr_scheduler_warmup_ratio: 0.1
|
| 42 |
+
- num_epochs: 3.0
|