README.md

---
license: mit
datasets:
- FreedomIntelligence/Huatuo26M-Lite
- CocoNutZENG/NeuroQABenchmark
language:
- zh
- en
metrics:
- accuracy
base_model:
- Qwen/Qwen2.5-7B-Instruct
tags:
- medical
pipeline_tag: text-generation
library_name: peft
---
## Introduction
Train LLM to be neuroscientist. It expected to work in Chinese and Engliah environment.

## Data
1. [FreedomIntelligence/Huatuo26M-Lite](https://huggingface.co/datasets/FreedomIntelligence/Huatuo26M-Lite). We select neuroscience(神经科学) label as train data.
2. [CocoNutZENG/NeuroQABenchmark](https://huggingface.co/datasets/CocoNutZENG/NeuroQABenchmark)


## Train Detail
We fine-tuned the Qwen2.5 model using supervised fine-tuning (SFT) with LoRA for efficient parameter optimization. The LoRA configuration employed a rank of 8 (R=8) to
balance adaptation quality with computational efficiency. Training was conducted for 1 epoch (approximately 1 hour duration) using two NVIDIA A40 GPUs with DeepSpeed’s Stage 2 optimization
for memory efficiency. We adopted the Adam optimizer with an initial learning rate of 5e-5 and a
cosine learning rate scheduler for smooth decay. This configuration achieved effective model adaptation while maintaining computational tractability on our hardware setup. Our model’s loss drop as
expected, see figure below for loss detail.
[![image.png](https://i.postimg.cc/d30KMJCQ/image.png)](https://postimg.cc/62FPnJzF)

## Evalution
| Model          | Acc   |
|----------------|-------|
| Qwen2.5-3b     | 0.788 |
| Qwen2.5-7b     | 0.820 |
| +HuatuoLite    | 0.832 |
| +Full data     | 0.848 |