modernbert-ragtruth-token-level-binary (Track B)

Fine-tuned answerdotai/ModernBERT-base, binary token-classification model for span-level RAG hallucination detection: given a (context, response) pair, labels every response token supported (0) or hallucinated (1), so it recovers character-level spans, not just a single per-response score.

This is the best-performing system in this project (F1 0.7631 at the response level) and the model deployed in the live demo (src/models/predict.py), superseding both Track A (hugoomezz/deberta-v3-ragtruth-hallucination) and Approach 1 (hugoomezz/modernbert-ragtruth-response-level). Its recipe matches LettuceDetect's approach: binary token labels, unweighted loss, and character-overlap span evaluation, with checkpoint selection on span-level F1 rather than response-level F1 (see ADR-020 below).

Intended use

Research / portfolio demonstration of span-level RAG hallucination detection on RAGTruth: highlighting exactly which characters of a generated response are unsupported by the retrieved context, not just a binary verdict. Powers this project's live demo (paste-your-own text, and a real RAG pipeline over a small Wikipedia corpus). Not validated outside RAGTruth's three task types (QA, Summary, Data2txt), and not intended for production moderation decisions without further evaluation on your own data and threshold.

Training data

RAGTruth (Niu et al., 2024, ACL, arXiv:2401.00396) — MIT-licensed, reproduced in this project's docs/THIRD_PARTY_LICENSES.md. 13,578 train / 1,512 val / 2,700 test rows (one more val row than Track A — see ADR-006/ ADR-011), tokenized at max_length=4096 (0% of rows truncated). Per-token binary labels: 0 = supported, 1 = hallucinated (any character overlap with a gold annotated span); context and special tokens are ignored in the loss (label -100). Plain cross-entropy, no class weighting — an earlier 3-class BIO scheme with inverse-frequency weighting on an ultra-rare class caused near-total span fragmentation (0.037 F1); this binary/unweighted redesign fixed it.

Metrics (RAGTruth test set, n=2700)

Response-level (a response is "predicted hallucinated" iff any response token is predicted positive) — reported figures from this project's unified cross-system evaluation, matching the deployed model and the README's comparison table:

This matches LettuceDetect-base's published example-level F1 of 76.07%, from results/finetuned_track_b_token_level_metrics.json.

Span-level (character-overlap, LettuceDetect's headline metric; from results/finetuned_track_b_token_level_metrics.json):

vs. LettuceDetect-base's published span-level F1 of 55.44%.

Per task_type (response-level derived):

Recipe correction (ADR-020): these are the weights from a controlled ablation's arm (b) — a faithful LettuceDetect-recipe replication (lr=1e-5, effective batch 8, 6 epochs, checkpoint selection on span-level F1 instead of response-level F1). The prior deployed weights (arm a) selected the checkpoint on response-level F1, which is structurally unable to distinguish tight spans from sloppy ones and was suppressing span-level performance; switching the selection metric alone (no architecture or data change) raised span-F1 from 0.5113 to 0.5321 (+2.1 points) and response precision from 0.7873 to 0.8359, at the cost of some recall (0.7381 → 0.7020). A companion arm testing Annotation-Confidence-Weighted Supervision (down-weighting the training loss on annotator-flagged "implicit_true" spans) did not clear its pre-registered bar and was rejected as a tested negative result — see the ADR for both arms' full numbers.

Limitations

• Summary remains the weakest task (F1 0.51, recall 0.436 — misses more than half of hallucinated summaries), though improved over Track A/Approach 1.
• Overconfident, not paranoid: misses 26.2% of hallucinated responses but false-alarms on only 10.7% of faithful ones (247 FN vs 188 FP in this project's error analysis) — the token-level decision rule (flag if any token crosses P≥0.5) structurally favors silence over alarm.
• "Subtle" hallucinations are the hardest case: responses annotated only with "Subtle" span types are missed 40.3% of the time (vs 27.0% for evident-only), and detection F1 drops to ≈0.48–0.52 on GPT-3.5/GPT-4 outputs specifically — the deployment scenario (strong generators, subtle errors) where detection matters most.
• Known false positive on close paraphrase: during live demo testing, the model flagged a factually correct grounded answer that paraphrased the source ("second of six children, five siblings") as unsupported at score 0.99 — plausibly triggered by surface-form sensitivity rather than a genuine factual disagreement. Noted as a known limitation, not further investigated (docs/notes.md, Phase 5 section). This is the inverse of the "overconfident" pattern above: mostly the model under-flags, but this shows it can occasionally over-flag on close paraphrases.
• The decision threshold is a product tradeoff, not a fixed answer: F1 is nearly flat across thresholds 0.2–0.7, so the same checkpoint can be run as a high-precision "block" mode (t=0.9: precision 0.879, recall 0.602) or a high-recall "warn" mode (t=0.1: recall 0.843, precision 0.642). Even at the most aggressive setting, ~16% of hallucinations slip through — a risk reducer, not a guarantee.

How to use

This snippet reproduces the response-level score (max P(hallucinated) over response tokens); see src/models/predict.py in the project repo for full character-span reconstruction (merge_predicted_spans).

import torch
from transformers import AutoModelForTokenClassification, AutoTokenizer

# The tokenizer is loaded from the base ModernBERT repo, not the fine-tuned one: this
# repo's tokenizer_config.json was written by a newer transformers than some pinned
# installs can parse. Training data was built with this exact base tokenizer, so the
# substitution is not just safe but exact.
TOKENIZER_ID = "answerdotai/ModernBERT-base"
MODEL_ID = "hugoomezz/modernbert-ragtruth-token-level-binary"
SUPPORTED, HALLUCINATED = 0, 1

tokenizer = AutoTokenizer.from_pretrained(TOKENIZER_ID)
model = AutoModelForTokenClassification.from_pretrained(MODEL_ID, attn_implementation="sdpa").eval()

context = "The Eiffel Tower was completed in 1889 for the World's Fair in Paris."
response = "The Eiffel Tower was completed in 1889 and is located in Berlin, Germany."

encoding = tokenizer(
    context, response, max_length=4096, truncation="only_first",
    return_offsets_mapping=True, return_token_type_ids=False, return_tensors="pt",
)
sequence_ids = encoding.sequence_ids(0)
encoding.pop("offset_mapping")  # only needed for span reconstruction, not the score

with torch.no_grad():
    logits = model(**encoding).logits[0]
probs_hallucinated = torch.softmax(logits, dim=-1)[:, HALLUCINATED].tolist()

# Response-level score: max P(hallucinated) over response tokens (sequence_id == 1).
response_probs = [p for p, sid in zip(probs_hallucinated, sequence_ids) if sid == 1]
score = max(response_probs) if response_probs else 0.0
print(f"hallucination score: {score:.4f}  ({'FLAGGED' if score >= 0.5 else 'clean'})")

Citation

@inproceedings{niu2024ragtruth,
  title     = {RAGTruth: A Hallucination Corpus for Developing and Evaluating RAG Systems},
  author    = {Niu, Cheng and Wu, Yuanhao and Zhu, Juno and Xu, Siliang and Shum, Kashun and Zhong, Randy and Song, Juntong and Zhang, Tong},
  booktitle = {Proceedings of ACL 2024},
  year      = {2024},
  eprint    = {2401.00396}
}
@article{kovacs2025lettucedetect,
  title   = {LettuceDetect: A Hallucination Detection Framework for RAG Applications},
  author  = {Kov{\'a}cs, {\'A}d{\'a}m and Bakos, Zsolt},
  journal = {arXiv preprint arXiv:2502.17125},
  year    = {2025}
}