---
license: mit
base_model: microsoft/deberta-v3-base
tags:
  - token-classification
  - hallucination-detection
  - tool-use
datasets:
  - drond0174/RAGTruth-Hallucinations
---

# Hallucination detection artifacts (ToolACE / RAGTruth-style)

Checkpoints and test predictions for span-level hallucination detection in tool-augmented answers.

## Contents

| Path | Description |
|------|-------------|
| `deberta_contradiction_tuned/` | Tool-aware DeBERTa fine-tuned on mixed train (contradiction oversample ×3) — **best run** |
| `deberta_mixed/` | Earlier/alternate DeBERTa mixed checkpoint (no contradiction oversampling) |
| `predictions/` | `mixed_test` span predictions (DeBERTa, LookBack, Lettuce) |
| `lookback/lookback_mixed_classifier.joblib` | Sklearn head for LookBackLens (TinyLlama features) |
| `lookback/lookback_mixed_train_features.npz` | Cached train attention features (~1.1 GB) |
| `lookback/lookback_mixed_val_features.npz` | Cached validation attention features (~164 MB) |

Dataset: [drond0174/RAGTruth-Hallucinations](https://huggingface.co/datasets/drond0174/RAGTruth-Hallucinations)

## Load DeBERTa

```python
from transformers import AutoModelForTokenClassification, AutoTokenizer

model_dir = "drond0174/hallucination_detection"
tokenizer = AutoTokenizer.from_pretrained(f"{model_dir}/deberta_contradiction_tuned")
model = AutoModelForTokenClassification.from_pretrained(
    f"{model_dir}/deberta_contradiction_tuned"
)
```

See `deberta_contradiction_tuned/run_meta.json` for threshold, best epoch, and validation F1.

## LookBack feature caches

Download `lookback/*_features.npz` to skip re-running TinyLlama feature extraction. Point `train_cache_path` / `val_cache_path` in `lookback_baseline.py` to the downloaded files.