dipta007/decomposeRL-7b
dipta007/decomposeRL-7b is a 7.6 billion parameter fact-verification model developed by Shubhashis Roy Dipta, Ankur Padia, and Francis Ferraro. Fine-tuned from Qwen2.5-7B-Instruct, it learns to decompose claims into sub-questions, iteratively answer them from provided evidence, and produce a 'Supported' or 'Refuted' judgment. This model excels at open-book fact verification, achieving high balanced accuracy across various claim-verification benchmarks, and provides fully transparent reasoning.
Loading preview...
Overview
DecomposeRL-7B is a 7.6 billion parameter fact-verification model, fine-tuned from Qwen2.5-7B-Instruct using GRPO + LoRA. Its core innovation lies in its ability to decompose complex claims into atomic sub-questions, iteratively answer them from an evidence document, and then provide a final Supported or Refuted judgment. The model's training incorporates a stack of seven complementary rewards designed to ensure structural correctness, per-question quality, and overall sufficiency of the verification process.
Key Capabilities
- Transparent Reasoning: Emits a full trace including sub-claim checklists, questions asked, evidence quotes, and the final label.
- High Accuracy: Achieves 84.4% micro-average balanced accuracy and 86.3% macro-average balanced accuracy across 9 in-domain claim-verification benchmarks.
- Robust on Long-Form Evidence: Demonstrates strong performance on datasets like Ex-FEVER (88%), FEVEROUS (93%), and HoVer (76%).
- Out-of-Domain Performance: Scores 60.2% balanced accuracy on Coverbench and 77.0% on LLM-AggreFact.
- Decompose-Question-Answer-Verify Loop: Follows a structured process involving initial analysis, iterative QA, sufficiency checks, and a final verdict.
Good for
- Open-book fact verification: Verifying factual claims against a provided evidence document.
- Retrieval-augmented fact-checking pipelines: Integrating into systems where evidence is retrieved and then verified.
Limitations
- Out-of-scope for closed-book fact-checking: Not intended for verification against the model's parametric knowledge or real-time news verification without supplied evidence. The model is trained to say "I don't know" if evidence is silent.