FaithLens
Properties
created
08.06.2026, 10:00
modified
26.07.2026, 13:41
published
25.12.2025, 00:00
topics
Grounded Fact Verification, Evaluation
authors
Opus 4.7
ai-assisted
Yes
- Resources
- Paper:
FaithLens (Si et al., 2025)
- ACL 2026
- GitHub: S1s-Z/FaithLens
- Model: ssz1111/FaithLens
- Paper:
FaithLens (Si et al., 2025)
- 8B specialized faithfulness classifier that, unlike Bespoke-MiniCheck-7B, also returns a natural-language justification alongside the label.
# What It Does
- Dual task per
(document, claim):- Predict whether the claim is faithful or hallucinated wrt the document
- Explain the prediction in human-readable text
- Authors report the 8B model outperforming larger general-purpose LLMs (GPT-5.2, o3) on 12 grounding benchmarks including LLM-AggreFact and HoVer.
# Training Recipe
- Two-stage pipeline:
- Cold-start SFT on synthetic
(doc, claim, label, explanation)data - Rule-based RL rewarding both prediction accuracy and explanation quality
- Cold-start SFT on synthetic
- The explanation channel is what differentiates FaithLens from MiniCheck-style binary classifiers - useful when downstream users need to know why a claim was flagged, not just that it was.
# Why It Matters
- For audit / compliance flows where a “0/1 hallucinated” verdict is not enough.
- For Claim Decomposition pipelines that want to surface the exact missing/contradicting evidence to the end user rather than just suppressing the answer.