katsubakirill/toolace-hallucination-span-detection
Preview • Updated • 17
Fine-tuned ModernBERT token classifier for span-level hallucination detection in tool-calling dialogues.
KRLabsOrg/lettucedect-base-modernbert-en-v1
katsubakirill/toolace-hallucination-span-detection
tool-aware input: question, available tools, executed tool call, tool facts, raw tool output, answer
Only answer tokens are trained/evaluated for hallucination labels. Non-answer
tokens are ignored with label -100.
0: non_hallucinated1: hallucinated| dataset_type | n_examples | span_precision | span_recall | span_f1 | example_f1 | exact_span_match_rate | threshold |
|---|---|---|---|---|---|---|---|
| overall | 330 | 0.795651 | 0.757896 | 0.776315 | 0.80678 | 0.506061 | 0.99 |
| contradiction_with_tool_output | 130 | 0.259182 | 0.31385 | 0.283908 | 0.571429 | 0.469231 | 0.99 |
| missing_tool | 62 | 0.964447 | 0.849658 | 0.903421 | 0.95082 | 0.629032 | 0.99 |
| overgeneration | 138 | 0.858434 | 0.802064 | 0.829292 | 0.911765 | 0.485507 | 0.99 |
Base model
answerdotai/ModernBERT-base