modernbert-ragtruth-response-level (Approach 1)

Fine-tuned answerdotai/ModernBERT-base, binary sequence classifier for response-level RAG hallucination detection: given a (context, response) pair, predicts whether the response contains any hallucinated content relative to the context. Label 0 = faithful, label 1 = hallucinated. Same task as Track A (hugoomezz/deberta-v3-ragtruth-hallucination), on a truncation-free backbone.

This is not the best-performing system in this project — see hugoomezz/modernbert-ragtruth-token-level-binary (Track B), which supersedes it and is the model deployed in the live demo.

Intended use

Research / portfolio demonstration of RAG hallucination detection on RAGTruth, and a long-context (ModernBERT, 4096 tokens) comparison point against Track A's DeBERTa-v3-base (512 tokens). Not validated outside RAGTruth's three task types (QA, Summary, Data2txt), and not intended for production moderation decisions without further evaluation.

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 response-level rows, class-weighted cross-entropy. Tokenized at max_length=4096, which eliminates truncation entirely on RAGTruth (0.00% of rows truncated, vs. 70.34% at DeBERTa-v3's 512-token limit).

Metrics (RAGTruth test set, n=2700; from results/finetuned_approach1_modernbert_metrics.json)

Metric	Value
Precision	0.6839
Recall	0.7731
F1	0.7257
Accuracy	0.7959

Per task_type:

Task	F1	Recall
Data2txt	0.8506	0.8705
QA	0.5924	0.6813
Summary	0.5088	0.5686

Note: this project's unified cross-system comparison table (README.md) reports F1 0.7254 for this model, a ~0.0004 discrepancy from the 0.7257 above traced to exactly one borderline test example flipping prediction between the original Kaggle training run and a later re-inference pass — consistent with ordinary GPU inference nondeterminism (batch composition / kernel selection), not a data or pipeline bug (see docs/notes.md).

Limitations

• Summary recall (0.5686) improved substantially over Track A (0.2451) by moving to a longer context window, but Summary F1 (0.5088) is still the weakest of the three task types.
• Precision is slightly lower than Track A's (0.6839 vs 0.7367) — the longer context window trades some precision for the large recall gain, concentrated almost entirely in Summary.
• Response-level only — cannot localize which part of a response is hallucinated. For span-level localization, use Track B instead.

How to use

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

MODEL_ID = "hugoomezz/modernbert-ragtruth-response-level"
tokenizer = AutoTokenizer.from_pretrained(MODEL_ID)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_ID).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 stands in Paris, France."

inputs = tokenizer(context, response, truncation="only_first", max_length=4096, return_tensors="pt")
with torch.no_grad():
    logits = model(**inputs).logits
probs = torch.softmax(logits, dim=-1)[0]
print({"faithful": probs[0].item(), "hallucinated": probs[1].item()})

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}
}