v1.3.0 — FactCG-DeBERTa-v3-Large Default
What's New
Default NLI Model: FactCG-DeBERTa-v3-Large
- 75.8% balanced accuracy on LLM-AggreFact (29K samples), up from 66.2%
- FactCG instruction template for correct 2-class inference
- SummaC-style source chunking with max-score aggregation
- Updated benchmark docs and competitor comparison
Full Changelog: v1.2.1...v1.3.0