QuantumCuddle/DistillAgent-PaperQA-3B
DistillAgent-PaperQA-3B by QuantumCuddle is a 3.1 billion parameter agentic QA model, distilled from Qwen2.5-3B-Instruct. It is specifically fine-tuned using LoRA/rsLoRA for question answering over scientific papers (QASPER) with constrained Thought/Action/Observation/Final Answer trajectories. This model excels at providing practical agentic behavior for research-paper QA, outperforming its base model in exact match and F1 scores on scientific paper question answering tasks.
Loading preview...
DistillAgent-PaperQA-3B: Compact Agentic QA for Scientific Papers
DistillAgent-PaperQA-3B is a 3.1 billion parameter model developed by QuantumCuddle, distilled from the Qwen/Qwen2.5-3B-Instruct base model. It is specifically fine-tuned using LoRA/rsLoRA with constrained agentic trajectories for question answering over scientific papers (QASPER).
Key Capabilities
- Agentic QA: Designed for practical agentic behavior when answering questions based on research papers.
- Performance: Outperforms its base model, Qwen2.5-3B-Instruct, in Exact Match (14.5% vs 9.0%) and Mean F1 (0.2425 vs 0.1650) on a 200-sample QASPER evaluation.
- Efficient: A compact 3.1B parameter model, making it suitable for scenarios where smaller models are preferred.
- Inference Style: Utilizes constrained ReAct with section lookup for answering questions.
Good For
- Scientific/Technical Paper QA: Ideal for question answering tasks that involve section-level lookup or retrieval within scientific or technical documents.
- Research & Education: Useful for exploring and implementing compact agentic model distillation in research and educational workflows.
Limitations
- Sensitive to the specific runtime prompt and harness format.
- Multi-hop reasoning can lead to increased latency.
- Not recommended as a sole source for high-stakes scientific or medical decisions.