orailix/MUCH-signals

[Signal-only dataset] MUCH: A Multilingual Claim Hallucination Benchmark

Jérémie Dentan¹, Alexi Canesse¹, Davide Buscaldi^{1, 2}, Aymen Shabou³, Sonia Vanier¹

¹LIX (École Polytechnique, IP Paris, CNSR), ²LIPN (Université Sorbonne Paris Nord), ³Crédit Agricole SA

Important Notice: Signal-Only

This dataset contains only the evaluation signals of the baselines evaluated on the MUCH benchmark. The full benchmark dataset is available at https://huggingface.co/datasets/jdentan/MUCH.

Dataset Description

Claim-level Uncertainty Quantification (UQ) is a promising approach to mitigate the lack of reliability in Large Language Models (LLMs). We introduce MUCH, the first claim-level UQ benchmark designed for fair and reproducible evaluation of future methods under realistic conditions. It includes 4,876 samples across four European languages (English, French, Spanish, and German) and four instruction-tuned open-weight LLMs. Unlike prior claim-level benchmarks, we release 24 generation logits per token, facilitating the development of future white-box methods without re-generating data. Moreover, in contrast to previous benchmarks that rely on manual or LLM-based segmentation, we propose a new deterministic algorithm capable of segmenting claims using as little as 0.1% of the LLM generation time. This makes our segmentation approach suitable for real-time monitoring of LLM outputs, ensuring that MUCH evaluates UQ methods under realistic deployment constraints. Finally, our evaluations show that current methods still have substantial room for improvement in both performance and efficiency.

Related artifacts

Alongside this dataset, we provide the following resources:

• Our research paper introducing MUCH and describing its generation in detail: arXiv:2511.17081
• A GitHub repository implementing the generation and utilization of MUCH: orailix/much
• The dataset, available on HuggingFace:
  • Main dataset: orailix/MUCH
  • Generation configs: orailix/MUCH-configs
  • Baseline evaluation data: orailix/MUCH-signals
  • "Trash" split containing the samples filtered-out from MUCH: orailix/MUCH-trash-only-for-reproducibility
• A PyPI package implementing our claim segmentation algorithm: much-segmenter

Baselines description

This dataset contains the evaluation token-level signal computed for the five baselines evaluated on MUCH: CCP [1], SAR [2], Max-likelihood [3], Likelihood [4] and Token Entropy [5].

CCP [1] First, we adapted CCP to MUCH. The main adaptation arises from the fact that the meaning of claims in MUCH is often not self-contained, whereas the original method operates on self-contained sentences. To preserve the performance of the Natural Language Inference (NLI) model used in CCP, we provide the model with the sigma = 8 preceding and succeeding tokens of the token for which the CCP score is computed. Moreover, we evaluated the delta = 24 possible alternatives for each token (the original paper used delta = 10).

SAR [2] Second, we adapted SAR to MUCH. Like CCP, it relies on an entailment score computed on sentences. For the same reason, we fed the NLI model with the sigma = 8 preceding and succeeding tokens around the target token.

Max-Likelihood [3] Third, we adapted Maximum Likelihood, where the token-level UQ score corresponds to the probability of the most likely token at each generation step.

Token Likelihood [4] Fourth, Token Likelihood is defined as the likelihood of the sampled token.

Token Entropy [5] Finally, Token Entropy (T-E) is the entropy of the probability distribution over the top-24 tokens provided in MUCH.

Dataset Structure

Data Fields

These scores are intended for the evaluation of baseline methods on the MUCH benchmark, using the scripts available in our GitHub repository:
https://github.com/orailix/much

generation_id: str The identifier of the sample for which the signal is computed.
This field can be cross-referenced with the generation_id field in the MUCH dataset.

signal_name: str
The name of the signal, including its hyperparameters when applicable.
- CCP: ccp_delta_sigma (hyperparameters: delta is in {10, 24}, sigma is in {3, 5, 8})
- SAR: sar_sigma (hyperparameter: sigma is in {3, 5, 8})
- Maximum Likelihood: max_likelihood (no hyperparameters)
- Token Likelihood: token_likelihood (no hyperparameters)
- Token Entropy: token_entropy_delta (hyperparameter: delta is in {5, 10, 24})

values: list[float] The signal values for each token in the generation.
Values are set to $0.0$ for stopwords (see details in our paper).

Dataset Creation

For details about the dataset design, annotation process, and generation pipeline, please refer to our paper.
All generation and preprocessing scripts are available at https://github.com/orailix/much.

Usage

This dataset provides signals intended for evaluation on the MUCH dataset, using the scripts available at
https://github.com/orailix/much.

Acknowledgement

The prompt, wiki_url, and lang fields of the MUCH samples are extracted from the Mu-SHROOM [6], a dataset released under CC-BY-4.0 license.

This work received financial support from the research chair Trustworthy and Responsible AI at École Polytechnique.

This work was granted access to the HPC resources of IDRIS under the allocation AD011014843R1, made by GENCI.

References

• [1] Ekaterina Fadeeva, Aleksandr Rubashevskii, Artem Shelmanov, Sergey Petrakov, Haonan Li, Hamdy Mubarak, Evgenii Tsymbalov, Gleb Kuzmin, et al. Fact-Checking the Output of Large Language Models via Token-Level Uncertainty Quantification Findings of the ACL pages 9367–9385. 2024. https://aclanthology.org/2024.findings-acl.558/
• [2] Jinhao Duan, Hao Cheng, Shiqi Wang, Alex Zavalny, Chenan Wang, Renjing Xu, Bhavya Kailkhura, and Kaidi Xu. Shifting Attention to Relevance: Towards the Predictive Uncertainty Quantification of Free-Form Large Language Models ACL, volume 1, pages 5050–5063. 2024. https://aclanthology.org/2024.acl-long.276/
• [3] Lukas Aichberger, Kajetan Schweighofer, and Sepp Hochreiter. Rethinking Uncertainty Estimation in Natural Language Generation QUESTION workshop at ICLR. 2025. https://openreview.net/forum?id=iBKWqXCSFA
• [4] Nuno M. Guerreiro, Elena Voita, and André Martins. Looking for a Needle in a Haystack: A Comprehensive Study of Hallucinations in Neural Machine Translation EACL, pages 1059–1075. 2023. https://aclanthology.org/2023.eacl-main.75/
• [5] Andrey Malinin and Mark Gales. Uncertainty Estimation in Autoregressive Structured Prediction ICLR. 2021. https://openreview.net/forum?id=jN5y-zb5Q7m
• [6] Raúl Vázquez, Timothee Mickus, Elaine Zosa, Teemu Vahtola, Jörg Tiedemann, Aman Sinha, Vincent Segonne, Fernando Sánchez-Vega, Alessandro Raganato, Jindřich Libovický, Jussi Karlgren, Shaoxiong Ji, Jindřich Helcl, Liane Guillou, Ona de Gibert, Jaione Bengoetxea, Joseph Attieh, Marianna Apidianaki. SemEval-2025 Task 3: Mu-SHROOM, the Multilingual Shared Task on Hallucinations and Related Observable Overgeneration Mistakes ArXiv preprint. 2025. https://arxiv.org/abs/2504.11975

Copyright and License

Copyright 2025–present Laboratoire d’Informatique de l’École Polytechnique.

This dataset is released under the Apache-2.0 license.

Please cite this dataset as follows:

@misc{dentan_much_2025,
  title = {MUCH: A Multilingual Claim Hallucination Benchmark},
  author = {Dentan, Jérémie and Canesse, Alexi and Buscaldi, Davide and Shabou, Aymen and Vanier, Sonia},
  year = {2025},
  url = {https://arxiv.org/abs/2511.17081},
}