🐱 FineCat-NLI: 8,192-Token RAG Claim & Hallucination Verifier
Powered by dleemiller/finecat-nli-l — ModernBERT-Large (395M)
Verify whether a generated sentence is Entailed, Neutral, or Contradicted by long retrieved contexts up to 8,192 tokens.
Pre-loaded Test Scenarios
| 📖 Premise (Retrieved Context / Evidence Document) | 💬 Hypothesis (Generated Claim / Assertion to Verify) |
|---|
Audit an entire multi-sentence AI answer against source evidence. Breaks down the answer into individual claims and evaluates Quality Gate pass/fail.
Claim ID | Sentence Statement | Verdict | Entailment | Neutral | Contradiction |
|---|---|---|---|---|---|
ModernBERT-Large Backbone
- Model:
dleemiller/finecat-nli-l - Parameters: 395M
- Context Window: 8,192 tokens (RoPE rotary position embeddings)
- Attention: FlashAttention-2 + Unpadding (efficient throughput on modern GPUs)
- VRAM Footprint: ~1.8 GB peak (43% reduction compared to DeBERTa-v3-large)
Training Methodology
- Dataset Synthesis: Concatenated 6 premier NLI benchmarks: MNLI, SNLI, ANLI, WANLI, LingNLI, and NLI-FEVER.
- Quality Downsampling: Excluded easy superficial examples (>0.9 teacher confidence) by ~60%, forcing the model to learn deep semantic reasoning rather than shallow keyword overlap.
- Logit Distillation: Distilled from state-of-the-art
MoritzLaurer/DeBERTa-v3-large-mnli-fever-anli-ling-wanli.