Update README.md
Browse files
README.md
CHANGED
|
@@ -10,7 +10,28 @@ base_model:
|
|
| 10 |
pipeline_tag: text-generation
|
| 11 |
---
|
| 12 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 14 |
|
| 15 |
|
| 16 |
|
|
|
|
| 10 |
pipeline_tag: text-generation
|
| 11 |
---
|
| 12 |
|
| 13 |
+
## Purpose
|
| 14 |
+
Helps you rewrite your voice-over script in the style of a popular Douyin (TikTok) creator. Simply put, 说人话！<br>
|
| 15 |
+
Currently, three distinct Douyin creator styles are used as references:<br>
|
| 16 |
+
"多多喂" – Known for exaggerated humor, high energy, and a down-to-earth, relatable tone.<br>
|
| 17 |
+
"Eyeopener" – A humorous science communicator with a lighthearted, vivid, and easy-to-understand approach.<br>
|
| 18 |
+
"严伯钧" – Another science-focused creator, but with a more straightforward and calm delivery."<br>
|
| 19 |
|
| 20 |
+
## Train
|
| 21 |
+
It's a reasoning model. Train Qwen/Qwen3-14B by USLOTH's GRPO with dataset (https://huggingface.co/datasets/aipgpt/douyin_style_v2)
|
| 22 |
+
|
| 23 |
+
|
| 24 |
+
## Deploy
|
| 25 |
+
vllm serve <model_path> --served-model-name <served_model_name> --dtype auto --kv-cache-dtype auto --gpu_memory_utilization 0.95 --host 0.0.0.0 --port 7000 --max_model_len 30000
|
| 26 |
+
|
| 27 |
+
|
| 28 |
+
## Test
|
| 29 |
+
Use streamlit style quick AI demo framework to write a testing program. <br>
|
| 30 |
+
prompt：<br>
|
| 31 |
+
[<br>
|
| 32 |
+
{"role": "system", "content": "请使用中文按以下格式回答问题:\n<think>\n...\n</think>\n\n..."},<br>
|
| 33 |
+
{"role": "user", "content": "你是抖音博主`<douyin_creator_name>`, 请把所给的文稿改写成`<douyin_creator_name>`的抖音口播稿文案，并用中文输出。你的文案从不用表情符号，也不包含拍摄指导脚本。n\n`<your_script>`"}<br>
|
| 34 |
+
]
|
| 35 |
|
| 36 |
|
| 37 |
|