phenixace/Chem-R-Faithful
Chem-R-Faithful is an 8 billion parameter language model developed by phenixace, derived from Chem-R-8B. It is specifically fine-tuned using a verification-grounded process reward (GRPO) to significantly reduce fabrication and improve the faithfulness of its reasoning traces in chemistry-related generative tasks. This model excels at producing chemically sound explanations by ensuring functional group claims are supported by input or predicted molecules, making it suitable for applications requiring high factual accuracy in chemical reasoning.
Loading preview...
Chem-R-Faithful: Enhanced Chemical Reasoning with Reduced Fabrication
Chem-R-Faithful is an 8 billion parameter model developed by phenixace, building upon the Chem-R-8B base. Its core innovation lies in its training methodology: it was fine-tuned using GRPO (Generative Reinforcement Learning with Policy Optimization) under a unique verification-grounded process reward. This reward mechanism ensures that accuracy is only granted when the model's reasoning trace is 'clean' – meaning every functional-group claim is verifiable against the input, predicted molecule, or reference.
Key Differentiators and Performance
Unlike models rewarded solely on answer correctness, Chem-R-Faithful's reward system directly penalizes fabricated reasoning steps. This approach leads to substantial improvements in faithfulness without sacrificing task performance. Across twelve generative chemistry tasks (including ChEBI-20 caption↔molecule, USPTO-50k retrosynthesis, and S²-Bench subtasks), Chem-R-Faithful demonstrates:
- A five-fold reduction in per-claim fabrication rate (from 22.56% to 4.35%).
- A doubling of the clean-trace rate (from 40.63% to 88.05%).
- A slight increase in overall task performance (from 50.09 to 51.45).
Usage and Output Format
The model outputs responses in a <think>…</think><answer>…</answer> format, allowing for separate parsing of the final answer and the reasoning trace. The internal detector audits the trace span for faithfulness. Training involved 936 steps on 4×H200 GPUs, utilizing a reward function that heavily weights accuracy, groundedness, and anti-hallucination, with accuracy gated by a zero fabrication score.
Limitations
While significantly reducing fabrication, it's important to note that the verifier focuses on structurally decidable claims (functional groups, ring systems). It does not certify a complete chemical argument, and fabrication is reduced, not entirely eliminated.