Nous ResearchMinos
Minos v1
Minos fine-tunes ModernBERT-large to classify a question and answer as refusal or non-refusal.
Training and model behavior
The classifier was trained on 387,134 examples and 132 million tokens. It expects explicit user and assistant delimiters. Class zero means non-refusal; class one means refusal. The output includes a predicted class and confidence.Minos v1 model card
Working with this model
Practical guidance. Use a refusal detector to measure response behavior. It does not assess factual accuracy, helpfulness or whether a refusal was appropriate. Review ambiguous and partial refusals separately.
Specifications & access
| Detail | Value and conditions |
|---|---|
| Launch date | April 2025 |
| Access | Downloadable |
| Model identifiers | NousResearch/Minos-v1 |
| Context window | 8,192 tokens Context specified by the model card. Minos v1 model card |
| Base model | answerdotai/ModernBERT-largeMinos v1 model card |
| Repository license label | apache-2.0Minos v1 model card |