---
base_model: unsloth/gpt-oss-20b-unsloth-bnb-4bit
tags:
- text-generation-inference
- transformers
- unsloth
- gpt_oss
- trl
license: apache-2.0
language:
- en
datasets:
- ericrisco/medrescue
---

# gpt-oss-20b–medical-rescue

**A \~20B parameter language model (base: `gpt-oss-20b`) fine-tuned with Unsloth on a curated 80k+ Q\&A dataset focused on first aid, emergency medicine, and clinical reasoning.**

⚠️ **IMPORTANT DISCLAIMER:** This model is for **educational and research purposes only**. It must **never** be used as a substitute for professional medical advice, diagnosis, or treatment. Always consult qualified healthcare professionals.

---

## Model Summary

* **Primary Task:** Medical and rescue Q\&A, procedural first aid guidance, and basic clinical reasoning.
* **Training Dataset:** [ericrisco/medical-training-dataset](https://huggingface.co/datasets/ericrisco/medical-training-dataset) (80k+ examples, CC-BY-4.0).
* **Fine-tuning Method:** Unsloth (PEFT).
* **Data Structure:**

  * `input`: medical question or scenario
  * `context`: additional information (optional)
  * `output`: detailed answer
  * `source`: original dataset or document reference

---

## Intended Use

**Suitable for:**

* Medical education and training simulations.
* First aid and emergency response learning tools.
* Research in medical NLP safety and reasoning.

**Not suitable for:**

* Real-time medical decision-making.
* Diagnosing conditions or prescribing treatments.
* Emergency triage or operational deployment in healthcare systems.

---

## Training Data

**Dataset:** [ericrisco/medical-training-dataset](https://huggingface.co/datasets/ericrisco/medical-training-dataset) (CC-BY-4.0).
**Size:** 80,000+ Q\&A pairs.

**Sources:**

* **11 public medical datasets** (e.g., *medical-o1-reasoning-SFT*, *medqa*, *medicationqa*, *symptom\_to\_diagnosis*).
* **14 official medical & emergency PDFs** from WHO, IFRC, FEMA, and governmental emergency guides.
* Synthetic Q\&A generation using **Ollama + Gemma3N** with a **RAG pipeline** powered by FAISS, Cohere embeddings, and strict medical relevance filtering.

**Processing Pipeline:**

1. **PDF extraction & chunking** with LangChain (512 tokens per chunk, 128 overlap).
2. **Semantic vectorization** via Cohere `embed-v4.0`.
3. **RAG-based question generation** from authoritative sources.
4. **Answer synthesis** with medical accuracy validation.
5. **Data standardization** into a unified schema.

---

## Fine-Tuning Details

* **Base model:** `gpt-oss-20b`
* **Framework:** [Unsloth](https://github.com/unslothai/unsloth) (LoRA/QLoRA PEFT).
* **Training objective:** Supervised fine-tuning (SFT) on high-quality Q\&A format.
* **Safety filters:** Removal of irrelevant, administrative, or unsafe content; enforced presence of key medical indicators (e.g., patient assessment, vital signs, treatment steps).
* **Dataset balance:** Shuffled and deduplicated to avoid overfitting.

---

## License

* **Model weights:** Same license as `gpt-oss-20b`.
* **Dataset:** CC-BY-4.0 — *Medical Training Dataset: Comprehensive Medical Q\&A for AI Training*.
* **Pipeline & scripts:** [GitHub – gemma3n-impact-challenge](https://github.com/ericrisco/gemma3n-impact-challenge).

---

## Citation

If you use this dataset or model, please cite:

```
@dataset{medical_training_dataset_2025,
  title={Medical Training Dataset: Comprehensive Medical Q&A for AI Training},
  author={Eric Risco},
  year={2025},
  url={https://huggingface.co/datasets/ericrisco/medical-training-dataset},
  repository={https://github.com/ericrisco/gemma3n-impact-challenge}
}

@misc{gpt_oss_20b_medical_rescue_2025,
  title={gpt-oss-20b–medical-rescue},
  author={Eric Risco},
  year={2025},
  howpublished={\url{https://huggingface.co/TU_USERNAME/gpt-oss-20b–medical-rescue}}
}
```