Project-Angel2
Collection
The collections is the new version of Project-Angel following the DeepSeek fine-tune training method with a refined dataset. • 6 items • Updated
Uploaded model
- Developed by: derek33125
- License: apache-2.0
- Finetuned from model : derek33125/PA-stage1-Qwen7B-300
This qwen2 model was trained 2x faster with Unsloth and Huggingface's TRL library.
Dataset
- EmoLLM
- SoulChat Multi-Trun Dataset
- Real mental health conversation data where the responses are replaced by DeepSeek-R1
Method
- We use GRPO fine-tuning to enhance its ability to think
- Reward model:
- Thinking length (prevent giving too short/long result)
- Calling other LLM with different setup to give the reward for the response (using average scoring for 3 judges)
- Answer Format
- Trained 147 steps
- Downloads last month
- 3
Model tree for derek33125/PA-stage2-Qwen7B-147
Base model
Qwen/Qwen2.5-7B Finetuned
Qwen/Qwen2.5-7B-Instruct Finetuned
unsloth/Qwen2.5-7B-Instruct Finetuned
derek33125/PA-stage1-Qwen7B-300