Create README.md
Browse files
README.md
ADDED
|
@@ -0,0 +1,58 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
datasets:
|
| 3 |
+
- IlyaGusev/ru_turbo_alpaca
|
| 4 |
+
- lksy/ru_instruct_gpt4
|
| 5 |
+
language:
|
| 6 |
+
- ru
|
| 7 |
+
pipeline_tag: text2text-generation
|
| 8 |
+
inference: false
|
| 9 |
+
---
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
Based on [LLaMA 30B](https://huggingface.co/yahma/llama-13b-hf).
|
| 13 |
+
|
| 14 |
+
Trained on 4 LoRA modules.
|
| 15 |
+
|
| 16 |
+
Parameters:
|
| 17 |
+
|
| 18 |
+
```
|
| 19 |
+
{
|
| 20 |
+
"base_model_name_or_path": "./llama-30b-hf",
|
| 21 |
+
"bias": "none",
|
| 22 |
+
"enable_lora": null,
|
| 23 |
+
"fan_in_fan_out": false,
|
| 24 |
+
"inference_mode": true,
|
| 25 |
+
"lora_alpha": 16,
|
| 26 |
+
"lora_dropout": 0.05,
|
| 27 |
+
"merge_weights": false,
|
| 28 |
+
"modules_to_save": null,
|
| 29 |
+
"peft_type": "LORA",
|
| 30 |
+
"r": 16,
|
| 31 |
+
"target_modules": [
|
| 32 |
+
"q_proj",
|
| 33 |
+
"v_proj",
|
| 34 |
+
"k_proj",
|
| 35 |
+
"o_proj"
|
| 36 |
+
],
|
| 37 |
+
"task_type": "CAUSAL_LM"
|
| 38 |
+
}
|
| 39 |
+
```
|
| 40 |
+
Cutoff length set to 512
|
| 41 |
+
|
| 42 |
+
|
| 43 |
+
```
|
| 44 |
+
Prompt template:
|
| 45 |
+
|
| 46 |
+
{
|
| 47 |
+
"description": "A shorter template to experiment with.",
|
| 48 |
+
"prompt_input": "### Задание:\n{instruction}\n\n### Вход:\n{input}\n\n### Ответ:\n",
|
| 49 |
+
"prompt_no_input": "### Задание:\n{instruction}\n\n### Ответ:\n",
|
| 50 |
+
"response_split": "### Ответ:"
|
| 51 |
+
}
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
[WandB report](https://wandb.ai/lksy/huggingface/runs/oj1ezptd)
|
| 55 |
+
|
| 56 |
+
Epochs: 4
|
| 57 |
+
|
| 58 |
+
Loss: 0.853
|