RAGHal-RAGTruth-auto-en-v1

Auto-annotated token-level hallucination training set for RAG faithfulness detection.

Built on top of RAGTruth train responses (original answers + prompts), with automatic span labels produced by our annotation pipeline.
No human span labels are used in this train set. Human annotations appear only in the official RAGTruth test split (for evaluation of models trained on this data).

This is the training corpus behind ZaandaTeika/RAGHal-large-en-v1.

Summary

Name	RAGHal-RAGTruth-auto-en-v1
Language	English
Split	train only (14 633 samples)
Tasks	QA, summarization, data-to-text
Label type	Character spans of unsupported / contradicting text in the answer
Answer source	Original RAGTruth generators (no re-generation)
Label source	Automatic (GPT-OSS-120B + critic + optional NLI)
Human labels	Not used for training

Dataset statistics

Task	Samples	With ≥1 hall. span	Hall. rate	# spans
QA	4 925	1 400	28.4%	2 471
Summary	4 758	1 778	37.4%	2 528
Data2txt	4 950	3 374	68.2%	7 269
Total	14 633	6 552	44.8%	12 268

Answers by generator (RAGTruth originals)

Generator	Samples	Hall. rate
gpt-3.5-turbo-0613	2 497	15.2%
gpt-4-0613	2 494	16.8%
mistral-7B-instruct	2 450	65.9%
llama-2-7b-chat	2 429	63.8%
llama-2-70b-chat	2 423	50.9%
llama-2-13b-chat	2 340	58.1%

Data fields

File: RAGHal-RAGTruth-auto-en-v1.jsonl (1 sample / line).

Field	Type	Description
query	string	Short task instruction / question
context	string	Source passages / document / structured JSON
output	string	Model answer (span offsets refer to this string)
task_type	string	"QA" | "Summary" | "Data2txt"
model	string	Answer generator (RAGTruth original LLM id)
hallucination_labels	list	Char spans in output; empty = clean
prompt	string	Full original RAGTruth prompt (for detector training)

Label example

{
  "start": 91,
  "end": 101,
  "text": " yesterday"
}

• start / end — character offsets into output
• text — output[start:end] (convenience duplicate of the span)

Empty hallucination_labels means the answer is fully supported by the source.

How this dataset was built

1. Base corpus (unchanged answers)

Item	Detail
Source	RAGTruth train
Sources × generators	2 515 × 6 = 15 090 responses
Generators	GPT-4, GPT-3.5-turbo, Mistral-7B-Instruct, Llama-2-7B/13B/70B-chat
Prompts	Original RAGTruth source_info templates (QA / Summary / Data2txt)
Re-generation	None — we annotate existing answers only

Task definitions:

Task	Source	What faithfulness means
QA	Question + retrieved passages (MS MARCO)	Answer must be supported by passages only
Summary	News document (CNN/DM)	Summary must not add facts beyond the document
Data2txt	Yelp JSON (business + reviews)	Text must stay within structured fields; null → silence

2. Automatic span annotation (no human train labels)

We label faithfulness (unsupported / contradicts SOURCE), not open-world factuality.
Annotator writes [HAL]…[/HAL] tags over hallucinated answer spans; tags are then converted to character offsets.

RAGTruth train responses (15 090)
              │
              ▼
     GPT-OSS-120B  Pass 1   (T=0.6)
       task-specific system prompt + few-shots
              │
              ▼
     GPT-OSS-120B  Critic   (T=0.3)
       removal-only: drop false-positive tags, never add new ones
              │
              ├─ Summary only ──► DeBERTa-large-MNLI (entailment thr=0.5)
              │                   drop spans entailed by the document
              ▼
     Postprocess
       · snap spans to word boundaries
       · Data2txt: merge adjacent spans
              │
              ▼
     Final train JSONL  (14 633 samples)

Stage	Tool / model	Role
Pass 1	openai/gpt-oss-120b via vLLM	Propose hallucinated spans
Critic	same model, stricter prompt	Remove over-predicted tags
NLI (Summary only)	microsoft/deberta-large-mnli	Drop spans entailed by the source
Runtime	multi-GPU vLLM	Batched annotation
Postprocess	custom rules	Snap / merge; discard failed rows

3. Task-specific annotation configs

Task	Prompt pack	Critic	NLI	Notes
QA	system + 5 human-gold few-shots (refusal / synthesis / contradiction)	yes	no	Final QA slice after QA re-annotation
Summary	system + few-shots	yes	yes (DeBERTa @ 0.5)	Kept from base auto-annot track
Data2txt	system + aligned Data2txt rules (null fields, subjective descriptors)	full re-annot (pass1 → critic)	no	Replaces older Data2txt labels

4. Final mix assembly

1. Start from base auto-annotated train set (QA + Summary + old Data2txt).
2. Keep QA and Summary slices.
3. Replace Data2txt with the aligned critic re-annotation.
4. After postprocess: 14 633 rows (348 Data2txt critic rows discarded).

No manual span editing on the train set.

Intended use

• Train token-level / span-level RAG hallucination detectors.
• Study automatic faithfulness labeling vs human RAGTruth test.
• Compare QA / Summary / Data2txt transfer within one annotation stack.

Not intended as gold human labels. Treat spans as silver / teacher annotations.

Evaluation protocol (recommended)

Train on this dataset; evaluate on official RAGTruth test (2 700) with human spans:

Metric	Definition
Example-level P/R/F1	Binary: any hallucinated span present
Span-level P/R/F1	Character-overlap precision / recall / F1

A model trained on this set: ZaandaTeika/RAGHal-large-en-v1.

Loading

data/final/RAGHal-RAGTruth-auto-en-v1.jsonl
data/final/README.md

from datasets import load_dataset

ds = load_dataset("json", data_files="data/final/RAGHal-RAGTruth-auto-en-v1.jsonl")
# after HF upload:
# ds = load_dataset("YOUR_ORG/RAGHal-RAGTruth-auto-en-v1")

row = ds["train"][0]
print(row["task_type"], row["model"], row["hallucination_labels"])
if row["hallucination_labels"]:
    span = row["hallucination_labels"][0]
    print(row["output"][span["start"]:span["end"]])

import json
from pathlib import Path

for line in Path("data/final/RAGHal-RAGTruth-auto-en-v1.jsonl").open(encoding="utf-8"):
    row = json.loads(line)
    ...

Limitations

• Labels are automatic (LLM teacher + critic + optional NLI); residual noise is expected.
• English only; RAGTruth task formats only.
• Hallucination rates differ strongly by generator (GPT-family ≪ open chat models).
• Data2txt is the noisiest / densest task (~68% positive).
• Not a substitute for human annotation when gold quality is required.

Citation

If you use this dataset, please cite RAGTruth:

@inproceedings{nie2024ragtruth,
  title={RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models},
  author={Nie, Fuxiang and Yao, Yufeng and Zhu, Jingheng and others},
  booktitle={ACL},
  year={2024},
}

Acknowledgements

Base responses and prompts: RAGTruth.
Annotation models: GPT-OSS-120B, DeBERTa-large-MNLI. Inference: vLLM.