OverwhelmingFit/SOPO-RL-RAGTruth10k

SOPO RL RAGTruth10k

This is the RAGTruth-based reinforcement-learning (RL) training data used by SOPO (Socratic Policy Optimization) for hallucination detection. Given source material and a generated response, the task is to identify spans in the response that are unsupported by or contradict the source.

Data

The repository contains one JSON Lines file, train_rl_10k_20251124_reasoning_areal.jsonl, with 10,090 records. It is used as the train split; no validation or test split is provided here. The records are derived from the training split of RAGTruth, which covers summarization, question answering, and data-to-text generation. This is an RL training dataset, not the RAGTruth evaluation set.

Key fields include:

• query, context, and output: the original instruction, source material, and generated response.
• task_type, model, quality, and temperature: task and source-response metadata.
• hallucination_labels: the source hallucination annotations, stored as a JSON-encoded string.
• prompt: chat messages for the hallucination-detection training task; input_str is a text representation of the input.
• reward_model.ground_truth.target: the reference answer for the training reward, stored as a JSON-encoded string with a hallucination list key.
• extra_info: fields used by the SOPO training code, including prompt, response, hallucination_labels, and task_type.

Some text fields include the original source passages and responses. For details of the original annotations, see the RAGTruth dataset documentation.

Usage

Load the dataset with datasets:

from datasets import load_dataset

dataset = load_dataset("OverwhelmingFit/SOPO-RL-RAGTruth10k", split="train")
print(len(dataset))  # 10090
print(dataset[0]["extra_info"].keys())

For the released SOPO training configuration, download the JSONL file and set data_path to its local path. The training code reads it with load_dataset(path="json", split="train", data_files=...) and uses extra_info for the prompt and reward computation. The same file is referenced by the released GRPO and CAPO configurations.

Provenance, rights, and limitations

This release reformats RAGTruth training examples for SOPO's RL training and reward interface. The upstream RAGTruth repository provides the original corpus and an MIT license; source passages and other third-party content may have separate terms. The SOPO disclaimer does not relicense third-party data. Review applicable upstream terms before reuse or redistribution.

Labels are annotations of generated responses, not guarantees about factual accuracy in other contexts. Do not use this training split as an independent evaluation benchmark.

Citation

If you use the original corpus, please cite RAGTruth:

@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}
}