OverwhelmingFit/SOPO-SFT-RAGTruth5k-Reasoning

SOPO SFT RAGTruth5k Reasoning

This is the reasoning-augmented supervised fine-tuning (SFT) training set for hallucination detection in SOPO. The task is to identify unsupported or contradicted spans in a generated response given its source material, and to return the spans as a JSON list.

The reasoning trajectories in this dataset were generated by Gemini 2.5 Flash. They are model-generated explanations for training, not human-written annotations or evidence that the model's intermediate claims are correct. The underlying source examples and hallucination annotations come from RAGTruth. The generated rationales were produced with access to the reference labels during data preparation; they should not be treated as independent assessments of the answers.

Data

The release contains train_sft_reasoning_5k_20251124.jsonl, the prepared SFT file with 4,952 examples. Although the dataset name says “5k,” this file contains 4,952 training records. Only a training split is provided; no validation or test data is included.

Each JSON Lines record contains:

• instruction: a one-element list containing the hallucination-detection prompt, source material, and candidate response.
• target: a one-element list containing a <think>...</think> reasoning section followed by a JSON object with a hallucination list key.

This is the prepared training file, not the original intermediate file of 5,000 candidate records with a separate gemini_thinking field. This dataset is distinct from the SOPO RL training data and the RAGTruth evaluation set. The associated SOPO cold-start model is released separately.

Usage

from datasets import load_dataset

dataset = load_dataset("OverwhelmingFit/SOPO-SFT-RAGTruth5k-Reasoning", split="train")
print(len(dataset))  # 4952
print(dataset[0]["instruction"][0])
print(dataset[0]["target"][0])

For SFT, use instruction[0] as the input and target[0] as the target text. The target includes both the reasoning section and the final hallucination-span list.

Provenance and limitations

The upstream RAGTruth repository provides the source corpus and an MIT license; source passages and other third-party materials may be subject to separate terms. The SOPO disclaimer does not relicense third-party data. Review all applicable terms before reuse or redistribution.

RAGTruth annotations and synthetic reasoning may be incomplete or inaccurate. Do not use this training split as an independent evaluation benchmark.

Citation

Please cite RAGTruth when using the underlying corpus:

@misc{wu2023ragtruth,
  title={RAGTruth: A Hallucination Corpus for Developing Trustworthy Retrieval-Augmented Language Models},
  author={Yuanhao Wu and Juno Zhu and Siliang Xu and Kashun Shum and Cheng Niu and Randy Zhong and Juntong Song and Tong Zhang},
  year={2023},
  eprint={2401.00396},
  archivePrefix={arXiv},
  primaryClass={cs.CL}
}