DeepPulse-Embedding Dense Retrieval Model Series
DeepPulse (深度把脉) is a member of 心语心言's open-source TCM series models. This project includes two dense retrieval (Embedding) models fine-tuned on different base models:
- DeepPulse-Embedding-m3: Fine-tuned based on
BGE-m3.
- DeepPulse-Embedding-0.6b: Fine-tuned based on
Qwen3-0.6B.
Both models are fine-tuned using a self-built medical dataset (especially TCM data) and are optimized for document chunk retrieval scenarios in the medical field.
Self-Built Dataset Evaluation Metrics
The evaluation results on the self-built medical dataset are as follows. It can be seen that the DeepPulse series models outperform the original base models on all metrics.
| Model Name | MRR | NDCG@10 | Recall@1 | Recall@5 | Recall@10 |
|---|
| Qwen3-0.6B | 0.9458 | 0.9566 | 0.9157 | 0.9822 | 0.99 |
| bge-m3 | 0.9418 | 0.9519 | 0.9109 | 0.98 | 0.9831 |
| DeepPulse-Embedding-0.6b | 0.9693 | 0.9751 | 0.9513 | 0.9891 | 0.9935 |
| DeepPulse-Embedding-m3 | 0.9729 | 0.9781 | 0.957 | 0.9896 | 0.9948 |
Acknowledgement
DeepPulse-Embedding was trained by the algorithm team from 心语心言.
- ChenCh2002 completed the code implementation, model training, and evaluation work.
- ChenCh2002 and Night-Quiet completed the data collection and organization work together.
- The project was led by Night-Quiet.
