burstiness-engine/sentence-length-dispersion

Sentence-Length Dispersion: Human Articles and Model Generations

The data behind Can Prompting Reliably Close the Sentence-Length Dispersion Gap? (Vittoria Lanzo and Dico Angelo, LP4FM workshop at NeurIPS 2026): the 200 human news articles every model is scored against, the generations of the nine benchmarked models, the DeepSeek V4 Pro leg re-run on 18 October 2026, the paraphrase probe run on the same snapshot, and the earlier paired ablation corpus. Scoring code, prompts, scored evaluations and the paper source are at https://github.com/VittoriaLanzo/burstiness-was-measured-wrong-lp4fm.

Contents

path	rows	contents
paper-dataset/prompts_200_human.jsonl	200	Wikinews headlines (2016–2021, CC BY 2.5) with their human articles; its id is the scenario_id in every generation file
<leg>-news-gen/<model>.jsonl	3,000 per leg (Qwen3.6-27B: 2,996; its four cells that exceeded the reasoning budget have no row)	generations of the nine benchmarked models: 200 headlines × 5 prompting modes × 3 samples
deepseek-v4-news-gen/benchmark-2026-10-18/deepseek-v4.jsonl	3,000	the DeepSeek V4 Pro grid re-run on 18 October 2026 through the API alias the July leg used, which then served a different checkpoint (documented that day as DeepSeek-V4-Pro-0813, a version dated 13 August, after the July run); DeepSeek V4 Pro† in the paper, reported beside the nine models and outside their counts
deepseek-v4-news-gen/paraphrase-2026-10-18/variant_v{1,2,3}.jsonl	200 each	the instruction-wording probe on the same snapshot: three paraphrases of the instruction prompt, one sample per headline
deepseek-v4-news-gen/*-2026-10-18/provenance/		the run records of each 18 October run: run manifest, run summary, repair and continuation records, the cell list (re-run) or job list (paraphrase probe), an appendix note, and the source dataset's README for the run
ablation-corpus/burstiness_pairs.jsonl	100	the earlier ablation corpus: 100 Wikinews articles (2016–2021, CC BY 2.5), each paired with one model-written article on the same headline (human_text, ai_text); its human side is the 100-article set on which clean_wikinews.py reproduces the shipped bodies exactly (Appendix D.1 of the paper)
MANIFEST.json		every file's SHA-256, size, source dataset, source revision, source path and transformation, and each data file's row count

Folders are named after the source datasets, except ablation-corpus, which holds burstiness-engine/burstiness-engine-ablation. Legs (one model's grid each): Qwen3.6-27B (qwen36), Gemma 3 27B (gemma3), DeepSeek V4 Pro (deepseek-v4), GLM-5.2 (glm52), Kimi K2.6 (kimi26), Grok 4.3 (grok43), Grok 4.5 (grok45), Gemini 3.1 Pro (gemini31), Claude Opus 4.8 (opus48).

Generation records

Each line is one generation. scenario_id joins it to the human article, eval_prompt_id is the prompting mode (control, instruction, fewshot, plan, self_refine; instruction_v1–instruction_v3 in the paraphrase probe), iteration is the sample (0–2) and text is the model's output as returned. The paper scores text after uniform normalization (experiments/ablation/normalize_generation.py in the repository), which regenerates the normalized text deterministically. An empty text marks a refusal or a failed cell; per-leg completeness is in Appendix D.2 of the paper. reasoning is empty in every generation file: reasoning text returned by a provider is not released.

How this dataset was assembled

Every file is copied from one of the project's private working datasets under burstiness-engine, at the revision recorded in MANIFEST.json. All files are byte-identical to their sources except the four 18 October generation files (the re-run grid and the three paraphrase variants), whose reasoning field is emptied; every other field and the row order are unchanged.

Not included: reasoning text, and the per-attempt logs, which contain it; normalized copies of the generations, which normalize_generation.py regenerates; evaluation JSONs and multiplicity tables computed before the segmenter correction described in the paper (the scored evaluations the paper reports are in the GitHub repository under experiments/results/); smoke-test runs; and the outputs that bench/scrub_thinking.py set aside because generation stopped inside the model's reasoning, whose cells were regenerated.

The generations name people as news articles do (public officials, quoted members of the public); these names are left as generated. The run records flag one retained off-topic output (DeepSeek V4 Pro†, self-refine, scenario 93, iteration 1).

Usage

from datasets import load_dataset
human = load_dataset("burstiness-engine/sentence-length-dispersion", "human_articles", split="train")
gens = load_dataset("burstiness-engine/sentence-length-dispersion", "kimi-k2.6", split="train")

Scoring one leg against the human articles, from the root of a clone of the repository, with $D a local download of this dataset (hf download burstiness-engine/sentence-length-dispersion --repo-type dataset --local-dir "$D"):

python experiments/ablation/normalize_generation.py $D/kimi26-news-gen/kimi-k2.6.jsonl kimi-k2.6.jsonl
python -m experiments.ablation.seed_avg_paired $D/paper-dataset/prompts_200_human.jsonl kimi-k2.6.jsonl kimi-k2.6

This writes seed_avg_paired_kimi-k2.6.json, byte-identical to experiments/results/benchmark/kimi-k2.6/seed_avg_paired.json in the repository. The repository's hf-datasets.zip is a copy of this dataset.

License

CC BY 2.5. The human articles are from English Wikinews (https://en.wikinews.org), by Wikinews contributors under CC BY 2.5, reduced to plain article bodies (the parse is described in Appendix D.1 of the paper).

Citation

@inproceedings{lanzo2026dispersion,
  title     = {Can Prompting Reliably Close the Sentence-Length Dispersion Gap?},
  author    = {Lanzo, Vittoria and Angelo, Dico},
  booktitle = {Workshop on Linguistic Principles for Foundation Models (LP4FM), NeurIPS 2026},
  year      = {2026}
}

Cite the data with the revision you used:

@misc{lanzo2026dataset,
  title        = {{Sentence-Length Dispersion: Human Articles and Model Generations}},
  author       = {Lanzo, Vittoria and Angelo, Dico},
  howpublished = {Hugging Face dataset},
  year         = {2026},
  url          = {https://huggingface.co/datasets/burstiness-engine/sentence-length-dispersion}
}