README.md

---
license: apache-2.0
tags:
  - 向量检索

#model-type:
##如 gpt、phi、llama、chatglm、baichuan 等
#- gpt

#domain:
##如 nlp、cv、audio、multi-modal
#- nlp

#language:
##语言代码列表 https://help.aliyun.com/document_detail/215387.html?spm=a2c4g.11186623.0.0.9f8d7467kni6Aa
#- cn

#metrics:
##如 CIDEr、Blue、ROUGE 等
#- CIDEr

#tags:
##各种自定义，包括 pretrained、fine-tuned、instruction-tuned、RL-tuned 等训练方法和其他
#- pretrained

#tools:
##如 vllm、fastchat、llamacpp、AdaSeq 等
#- vllm
  - 中医
  - 医疗
tasks:
  - sentence-embedding
base_model:
  - Qwen/Qwen3-Embedding-0.6B
base_model_relation: finetune
frameworks: PyTorch
language:
  - zh
---
# 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 |
| <font color="red">DeepPulse-Embedding-0.6b</font> | <font color="red">0.9693</font> | <font color="red">0.9751</font> | <font color="red">0.9513</font> | <font color="red">0.9891</font> | <font color="red">0.9935</font> |
| <font color="red">DeepPulse-Embedding-m3</font> | <font color="red">0.9729</font> | <font color="red">0.9781</font> | <font color="red">0.957</font> | <font color="red">0.9896</font> | <font color="red">0.9948</font> |

# Acknowledgement

DeepPulse-Embedding was trained by the algorithm team from 心语心言.

* [ChenCh2002](https://www.modelscope.cn/profile/ChenCh2002) completed the code implementation, model training, and evaluation work.
* [ChenCh2002](https://www.modelscope.cn/profile/ChenCh2002) and [Night-Quiet](https://huggingface.co/Night-Quiet) completed the data collection and organization work together.
* The project was led by [Night-Quiet](https://huggingface.co/Night-Quiet).