Grounded Fact Verification

• A branch of research on specialized models that decide whether a claim is supported by a grounding document/fact. The task is essentially classification:
  • supported - the claim is entailed by the document
  • not supported / orthogonal - the document does not contain enough evidence
  • contradicting - the document directly refutes the claim
• Historically this was the domain of Natural Language Inference (NLI) models trained on SNLI / MNLI / ANLI. Modern variants are purpose-built grounding classifiers (often 0.4B–8B) that drastically outperform NLI baselines on long, noisy documents typical of RAG outputs.
• Why this matters for production: these models are small enough and fast enough (often hundreds of docs/min on a single GPU) to be used as an online faithfulness metric, not just an offline eval. Pair them with Claim Decomposition to score each atomic claim in an LLM answer against retrieved context.

# Topics

• LLM-AggreFact
  • Aggregated benchmark of 11 grounded-factuality datasets; the de-facto leaderboard for this task.
• Bespoke-MiniCheck-7B
  • SOTA 7B grounding classifier (internlm2_5-7b-chat base) trained on ANLI + synthetic data.
• FaithLens
  • ACL'26 8B model that outputs both a faithfulness label and a human-readable explanation; trained with cold-start SFT + rule-based RL.
• Natural Language Inference
  • Classical predecessor task (entailment / neutral / contradiction); models like DeBERTa-MNLI and ANLI baselines are still used as cheap fallbacks.

# Leaderboard

## Production Use Case - Because these classifiers return a single `{0, 1}` label plus a probability, they can be wired in as a **streaming faithfulness gate**: - Decompose model answer into atomic claims (Claim Decomposition) - For each claim, run `MiniCheck(retrieved_doc, claim) -> prob` - Aggregate to a faithfulness score, block / flag responses below threshold - Throughput reference: Bespoke-MiniCheck-7B with vLLM hits **>500 docs/min on a single A6000** ([Bespoke-MiniCheck-7B HF card](https://huggingface.co/bespokelabs/Bespoke-MiniCheck-7B)).