deberta-v3-ragtruth-hallucination (Track A)

Fine-tuned microsoft/deberta-v3-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.

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. Track A is kept published for the project's 4-system comparison.

Intended use

Research / portfolio demonstration of RAG hallucination detection on RAGTruth. Given a retrieved context and a generated response, scores whether the response is faithful to that context. Not validated on any domain outside RAGTruth's three task types (QA, Summary, Data2txt), and not intended for production moderation decisions without further evaluation on your own data.

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 / 1,511 val / 2,700 test response-level rows (~55/45 faithful/hallucinated split in train), class-weighted cross-entropy to counter the imbalance. Context-only truncation at 512 tokens (ADR-004): the response is always fully preserved, only the context's head is truncated when the combined length exceeds the budget (this affects 70.34% of rows at DeBERTa-v3's 512-token limit).

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

Per task_type:

Limitations

• Summary is a significant weak spot (recall 0.245 — misses 3 out of 4 hallucinated summaries), the inverse of the zero-shot baseline's near-perfect recall.
• Context truncation affects 70.34% of RAGTruth rows at 512 tokens; its cost is precision-driven (over-flags faithful responses), not recall-driven, per this project's ADR-010 diagnostic.
• Response-level only — flags a response as hallucinated or not, but cannot localize which part of the response is hallucinated. For span-level localization, use Track B instead.

How to use

import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer

MODEL_ID = "hugoomezz/deberta-v3-ragtruth-hallucination"
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=512, 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}
}