---
license: agpl-3.0
language: en
tags:
  - hallucination-detection
  - nli
  - deberta-v3
  - director-ai
datasets:
  - pminervini/HaluEval
  - pietrolesci/nli_fever
  - tals/vitaminc
  - anli
base_model: MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli-ling-wanli
pipeline_tag: text-classification
---

# DeBERTa-v3-base -- Hallucination Detection

Fine-tuned for hallucination detection as part of
[Director-AI](https://github.com/anulum/director-ai).

Smaller variant (184M params) of the large model. Lower accuracy but
faster inference — suitable for CPU deployments.

## Training

- **Base**: MoritzLaurer/DeBERTa-v3-base-mnli-fever-anli-ling-wanli
- **Data**: ~100K examples from HaluEval, FEVER, VitaminC, ANLI R3
- **Epochs**: 3, lr 2e-5, batch 32 (effective), class-weighted CE loss
- **Labels**: 0 = entailment, 1 = neutral, 2 = contradiction

## Usage

```python
from director_ai.core import NLIScorer

scorer = NLIScorer(model_name="anulum/deberta-v3-base-hallucination")
score = scorer.score("The capital of France is Paris.", "Paris is in Germany.")
```

## License

AGPL-3.0 | Commercial licensing: [anulum.li](https://www.anulum.li)