---
title: RAFT-QA
sdk: gradio
emoji: 💻
colorFrom: purple
colorTo: gray
pinned: true
thumbnail: >-
  https://cdn-uploads.huggingface.co/production/uploads/67c714e90b99a2332e310979/TrWTI_vofjfGgx4PILFY3.jpeg
short_description: Retrieval-Augmented Fine-Tuning for Question Answering
sdk_version: 5.25.2
language:
- en
tags:
- retrieval-augmented-learning
- question-answering
- fine-tuning
- transformers
- llm
license: mit
datasets:
- pubmedqa
- hotpotqa
- gorilla
library_name: transformers
model-index:
- name: RAFT-QA
  results:
  - task:
      type: question-answering
      name: Open-Domain Question Answering
    dataset:
      name: PubMedQA
      type: question-answering
    metrics:
    - name: Exact Match (EM)
      type: exact_match
      value: 79.3
    - name: F1 Score
      type: f1
      value: 87.1
---



# RAFT-QA: Retrieval-Augmented Fine-Tuning for Question Answering

## Model Overview

RAFT-QA is a sophisticated **retrieval-augmented** question-answering model designed to significantly enhance answer accuracy through the integration of **retrieved documents** during the fine-tuning process. By utilizing retrieval-enhanced training, it advances traditional fine-tuning techniques.

## Model Details

- **Base Model Options:** `mistral-7b`, `falcon-40b-instruct`, or other leading large language models (LLMs)

- **Fine-Tuning Technique:** RAFT (Retrieval-Augmented Fine-Tuning)

- **Retrieval Strategy:** FAISS-based document embedding retrieval

- **Training Datasets Included:** PubMedQA, HotpotQA, Gorilla

## How It Works

1. **Retrieve Relevant Documents:** FAISS efficiently retrieves the most pertinent documents in response to a query.

2. **Augment Input with Retrieved Context:** Incorporates the retrieved documents into the input data.

3. **Fine-Tune the Model:** The model learns to effectively weigh the retrieved context to produce improved answers.

## Performance Comparison

| Model                  | Exact Match (EM) | F1 Score |
|------------------------|------------------|----------|
| GPT-3.5 (baseline)     | 74.8             | 84.5     |
| Standard Fine-Tuning   | 76.2             | 85.6     |
| **RAFT-QA (ours)**     | **79.3**         | **87.1** |

## Usage

To load the model using the `transformers` library:

```python
from transformers import AutoModelForQuestionAnswering, AutoTokenizer

model_name = "your-hf-username/raft-qa"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForQuestionAnswering.from_pretrained(model_name)
```

## Limitations

- The model's performance is contingent on the quality of the retrieved documents.
- For optimal results, domain-specific tuning may be necessary.

## Citation

If you utilize this model in your work, please cite it as follows:

```
@article{raft2025,
  title={Retrieval-Augmented Fine-Tuning (RAFT) for Enhanced Question Answering},
  author={Your Name et al.},
  journal={ArXiv},
  year={2025}
}
```

## License

This model is released under the Apache 2.0 License.

---

This version provides clarity and conciseness, ensuring all sections are clear and correctly formatted according to the Hugging Face repository standards. Make sure the dataset type (`question-answering`) matches your intended use case.