Yunnnuy/factuality-rmbench-style

Factuality RM-Bench Style

Factuality RM-Bench Style is a controlled English dataset for studying whether reward models and representation probes prefer stylistic presentation over factual correctness. Each row contains one question, a localized correct and incorrect proposition, and six responses formed by crossing correctness with three presentation styles: concise, normal, and Markdown.

This repository is an export package for factuality_rmbench_style_v6. The published data contain accepted groups only. Raw API responses, verifier explanations, rejected records, credentials, and internal service metadata are intentionally excluded.

Configurations

All configurations expose a single train split. They are alternative views of the same accepted pool, not independent train, validation, and test splits. balanced and city_filled are verified by the exporter to be ID subsets of full.

Source Data

The construction starts from true claims in the public SAPLMA-style dataset mirror notrichardren/azaria-mitchell. Six source topics are retained:

• —cities
• —inventions
• —elements
• —animals
• —companies
• —scientific facts

The source consists of individually labeled factual statements rather than question-answer pairs. Only normalized, deduplicated statements labeled true are used as source claims. The source claim identifier and text are retained for traceability.

Before making this dataset repository public, the uploader must verify the source dataset's current license and redistribution terms and add the final license metadata. This package deliberately does not assert a license that was not recorded in the generation manifest.

Construction Process

1. 1.True-claim selection. The pipeline filters SAPLMA records to the six supported topics, keeps only true claims, normalizes topic aliases, and removes duplicate (topic, claim) entries.
2. 2.Fact extraction. An API model interprets the source statement as a structured content package containing an entity, relation, correct value, value type, and a question. These fields are not extracted by a fixed parser; they are model-produced and subsequently validated.
3. 3.Localized false counterpart. The model proposes an incorrect value for the same entity and relation. The canonical incorrect statement is formed by replacing the unique correct value in the source claim, rather than freely rewriting the entire statement.
4. 4.Canonical verification. A verifier checks that the source is true, the counterpart is false, both answer the same question, the entity and relation are unchanged, the difference is localized, and no unrelated claim is introduced.
5. 5.Detailed Markdown generation. The model renders parallel correct and incorrect Markdown responses from the same content package. Multiple candidates may be generated, after which deterministic checks select a structurally valid and length-matched pair.
6. 6.Markdown-to-Normal transformation. Markdown formatting is removed while preserving the same factual units and a detailed plain-text presentation.
7. 7.Normal-to-Concise transformation. Each normal response is shortened to the core proposition while preserving its assigned factual value.
8. 8.Style-chain verification. A verifier checks semantic preservation across styles, parallel correct/incorrect structure, absence of new independent facts, and absence of explicit correctness-label leakage.
9. 9.Deterministic filtering. Final validators check required fields, expected entity/relation/value coverage, localized value separation, Markdown structure, lack of Markdown in plain styles, correct/incorrect length balance, and the intended concise-to-normal-to-Markdown length ordering.

All released rows passed the canonical, style-semantic, structural, and length checks used by the v6 pipeline. This means they passed the recorded automated procedure; it does not constitute a guarantee that every generated factual judgment is error-free.

Data Structure

One row is one prompt group. Correct and incorrect variants remain together so that pairwise objectives can be constructed without regrouping by a potentially ambiguous external key.

The principal fields are:

All required text fields must already be strings in the source artifacts. source_id is the sole exception: a numeric source identifier is deliberately normalized to a string so the public schema remains stable across source mirrors.

schema.json gives the machine-readable field definitions. quality_report.json reports direct counts and length statistics for the exported files, plus aggregate rejection counts from the source artifacts.

Loading

from datasets import load_dataset

full = load_dataset("<namespace>/factuality-rmbench-style", "full")
balanced = load_dataset("<namespace>/factuality-rmbench-style", "balanced")
city_filled = load_dataset(
    "<namespace>/factuality-rmbench-style",
    "city_filled",
)

For pairwise probe training, select one style and use the corresponding correct_<style> and incorrect_<style> columns. For cross-style conflict analysis, compare variants from different style columns while retaining the same row ID.

Intended Uses

• —Training or analyzing reward-model probes under controlled factuality and style variation.
• —Constructing within-style correctness pairs and cross-style conflict pairs.
• —Measuring sensitivity to concise, plain detailed, and Markdown presentation.
• —Studying correction-preservation trade-offs in reward-model interventions.

The dataset is not designed as a general-purpose factual QA benchmark, a source of current real-world facts, or a substitute for human factual verification.

Limitations

• —Incorrect values and structured fields are model-generated and verifier judgments are also model-based. Automated agreement can miss subtle errors.
• —The source topics and templates are not representative of all factual reasoning domains.
• —full is topic-imbalanced, and city_filled intentionally contains more city examples than a balanced sample.
• —Style variants preserve a small factual content package through controlled elaboration and redundancy. They do not reproduce every form of naturally occurring writing style.
• —Factual statements can become outdated or depend on the source's conventions.
• —Users should preserve prompt-group boundaries when splitting the data to avoid leakage among variants of the same factual proposition.

Quality and Privacy Artifacts

The exporter computes all released-row counts directly from the JSONL files. It fails on missing required fields, non-string responses, duplicate IDs, or a derived configuration containing an ID absent from full. Only aggregate rejection counts are copied into quality_report.json; rejected examples and verifier natural-language reasoning are not published.

Citation and Attribution

Please cite the original SAPLMA/Azaria--Mitchell work and the paper introducing this controlled dataset. Replace the placeholders below with the final BibTeX entries before public release.

@inproceedings{azaria2023internal,
  title={The internal state of an LLM knows when it’s lying},
  author={Azaria, Amos and Mitchell, Tom},
  booktitle={Findings of the Association for Computational Linguistics: EMNLP 2023},
  pages={967--976},
  year={2023}
}

@misc{cho2026style,
  title         = {When Style Overrides Correctness: Post-Hoc Correctness Strengthening in Reward Models},
  author        = {Cho, Changhee and Yun, DongHyun and Park, Seoyoung and Lee, Jay-Yoon},
  year          = {2026},
  note          = {Manuscript},
  howpublished  = {\url{https://huggingface.co/datasets/Yunnnuy/factuality-rmbench-style}}
}