Spaces:
Running
Running
| title: Agri-Critique Research Assistant | |
| emoji: 🌾 | |
| colorFrom: green | |
| colorTo: yellow | |
| sdk: gradio | |
| sdk_version: 4.44.0 | |
| app_file: app.py | |
| pinned: false | |
| license: mit | |
| # Agri-Critique: Self-Correcting Agricultural QA System | |
| ## Overview | |
| **Agri-Critique** is an evidence-based question-answering system for agricultural research, powered by a fine-tuned Llama 3.1 8B model trained on 2,859 QA pairs with self-correction capabilities. | |
| ### Key Features | |
| - 🎯 **Evidence-Grounded Answers**: All responses cite scientific papers | |
| - 🔍 **Hallucination Detection**: Trained to identify and correct errors | |
| - 🔎 **Critique Mode**: Validate any answer for accuracy and errors | |
| - 💬 **Conversational**: Maintains context across multiple questions | |
| - 📚 **Citation-Aware**: Provides paper references for verification | |
| ### Two Interaction Modes | |
| 1. **💬 Q&A Mode**: Ask questions and get evidence-based answers | |
| 2. **🔍 Critique Mode**: Submit a question + proposed answer to validate for: | |
| - Hallucinations (fake findings) | |
| - Unsupported claims | |
| - Temporal/causal errors | |
| - Missing citations | |
| ### Model Details | |
| - **Base Model**: Llama 3.1 8B Instruct | |
| - **Fine-tuning**: LoRA (rank=32, alpha=64) | |
| - **Training Data**: | |
| - 1,814 standard QA pairs (63.4%) | |
| - 1,045 adversarial critique pairs (36.6%) | |
| - **Training Objective**: Mixed curriculum (answer + critique) | |
| ### Dataset Composition | |
| The model was trained on **Agri-Critique**, a novel dataset combining: | |
| 1. **Standard QA**: Evidence-based agricultural questions | |
| 2. **Adversarial QA**: Intentionally flawed answers with critiques | |
| - Hallucinations (fake findings) | |
| - Temporal mismatches | |
| - Causal reversals | |
| - Unsupported extrapolations | |
| ### Research Contribution | |
| This system demonstrates: | |
| - **Self-Correction**: Models can learn to detect their own errors | |
| - **Citation Integrity**: Reduces "fake citation" problem in RAG systems | |
| - **Domain Adaptation**: Specialized for agricultural/scientific QA | |
| ### Usage | |
| 1. Type your agricultural research question | |
| 2. View the evidence-based answer with citations | |
| 3. Expand "View Evidence Sources" to see paper excerpts | |
| 4. Ask follow-up questions for multi-turn conversation | |
| ### Example Questions | |
| - "What is the effect of nitrogen fertilizer on rice yield?" | |
| - "How does drought stress affect wheat production?" | |
| - "What are the benefits of crop rotation?" | |
| - "Explain the impact of climate change on agriculture." | |
| ### Limitations | |
| - **Domain-Specific**: Optimized for agricultural/scientific questions | |
| - **Evidence-Dependent**: Answers limited to indexed papers | |
| - **Not Hallucination-Free**: While reduced, some errors may occur | |
| - **English Only**: Currently supports English language only | |
| ### Citation | |
| If you use this system in your research, please cite: | |
| ```bibtex | |
| @article{agri-critique-2025, | |
| title={Agri-Critique: Training Self-Correcting Language Models for Agricultural Question Answering}, | |
| author={Your Name}, | |
| journal={arXiv preprint}, | |
| year={2025} | |
| } | |
| ``` | |
| ### Links | |
| - 📄 [Paper](https://arxiv.org/your-paper) (Coming soon) | |
| - 💻 [GitHub](https://github.com/your-repo) | |
| - 🤗 [Model](https://huggingface.co/your-username/agri-critique-llama) | |
| - 📊 [Dataset](https://huggingface.co/datasets/your-username/agri-critique) | |
| ### License | |
| MIT License - See LICENSE file for details | |
| ### Acknowledgments | |
| Built with: | |
| - Llama 3.1 by Meta | |
| - Hugging Face Transformers | |
| - FAISS for vector search | |
| - Streamlit for UI | |