README.md

---
license: cc-by-4.0
---

# 🧬 Thoth-mini

**Thoth-mini** is a lightweight version of Thoth, designed for **efficient and scalable biological protocol generation** while retaining strong scientific reasoning ability.

- 📄 **Paper**: *Unleashing Scientific Reasoning for Bio-experimental Protocol Generation via Structured Component-based Reward Mechanism* (ICLR 2026)
- 🔗 **GitHub**: https://github.com/manglu097/Thoth
- 🤗 **Dataset**: https://huggingface.co/datasets/manglu3935/SciRecipe

---

## 🔍 Model Overview

- **Base model**: Qwen3-4B  
- **Parameters**: 4B  
- **GPU memory**: ~8GB  
- **Primary task**: Biological experimental protocol generation  

Thoth-mini is trained with the same **Sketch-and-Fill paradigm** and **SCORE reward mechanism** as Thoth, offering a strong performance–efficiency trade-off.

---

## 🧠 Output Format

```
<think>  reasoning and planning </think>
<key>    structured machine-readable steps </key>
<orc>    natural language protocol </orc>
<note>   optional safety notes </note>
```

---

## 🚀 Usage

```python
from transformers import AutoModelForCausalLM, AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("manglu3935/Thoth-mini")
model = AutoModelForCausalLM.from_pretrained("manglu3935/Thoth-mini")
```

---

## ⚠️ Intended Use

For fast scientific reasoning experiments and scalable research deployment.  
Generated protocols must be reviewed by qualified experts prior to laboratory execution.

---

## 📖 Citation

```bibtex
@article{sun2025unleashing,
  title={Unleashing Scientific Reasoning for Bio-experimental Protocol Generation via Structured Component-based Reward Mechanism},
  author={Sun, Haoran and Jiang, Yankai and Tang, Zhenyu and others},
  journal={arXiv preprint arXiv:2510.15600},
  year={2025}
}
```