lettucedect-v2-taxonomy-head

LettuceDetect goes agentic — span-level hallucination detection across RAG, code, and tool output.

lettucedect-v2-taxonomy-head: Span Typing Head (encoder cascade)

Overview

lettucedect-v2-taxonomy-head types a hallucinated span — it does not find spans. It is a label-conditioned mmBERT-base bi-encoder that, given a span a binary detector already located, assigns a hallucination category and subcategory by embedding the span and taking the nearest taxonomy-label description (cosine). Paired with the binary encoder lettucedect-v2-mmbert-base, it forms a fully-encoder typed detector — detection + typing at encoder cost, no generative model.

• Taxonomy: 3 categories (contradiction, fabricated_reference, unsupported_addition) × 13 subcategories.
• Stage-B only: run a binary detector (e.g. lettucedect-v2-mmbert-base) first, then this head types each span.
• For detection and typing in a single pass, see the generative lettucedect-v2-qwen-2b.

Usage (cascade, via lettucedetect)

from lettucedetect.models.inference import HallucinationDetector

det = HallucinationDetector(
    method="transformer",
    model_path="KRLabsOrg/lettucedect-v2-mmbert-base",      # binary detector (finds spans)
    taxonomy_head="KRLabsOrg/lettucedect-v2-taxonomy-head",  # this head (types them)
)
spans = det.predict(context=[context], question=question, answer=answer, output_format="spans")
# [{"start": ..., "end": ..., "text": "...", "category": "contradiction", "subcategory": "numerical"}]

Training

A label-conditioned bi-encoder. A shared mmBERT-base encoder embeds (i) a span — the mean-pooled answer tokens of the input answer [CTX] context — and (ii) each taxonomy label as the mean-pooled embedding of its "name: description" text. Training minimizes cross-entropy between the span embedding and the label embeddings (temperature-scaled cosine similarity), jointly over the category and subcategory label sets, so a span is typed by its nearest label description. Because labels enter only as text, the same head scores any label in the taxonomy.

Trained on the per-span category/subcategory annotations of the unified code + prose hallucination benchmark — SWE-bench coding-agent traces, developer tool output, ACL / README / Wikipedia, plus RAGTruth and 14-language PsiloQA.

Performance

• Typing accuracy (given a gold span): category 0.82 / subcategory 0.64 (validation).
• End-to-end cascade (binary detector → this head), char-overlap typed-F1 on the unified test set: 0.461 (subcategory-gated 0.315). The generative lettucedect-v2-qwen-2b types in a single pass and is higher (typed-F1 0.585 / 0.468); this cascade is the option when you want typed spans from a small, fast encoder-only stack.

Citing

@misc{kovács2026documentgroundingspanlevelhallucination,
      title={Beyond Document Grounding: Span-Level Hallucination Detection over Code, Tool Output, and Documents}, 
      author={Ádám Kovács and Bowei He and Xue Liu and István Boros and Szilveszter Tóth and Gábor Recski},
      year={2026},
      eprint={2607.00895},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2607.00895}, 
}