amozo/character-name-concentration

Dataset: LLM-generated stories and character extractions

44,782 short stories generated by 20 LLMs under 8 mitigation strategies (plus supplementary runs), with 88,919 extracted characters. Stories were generated between 2026-05 and 2026-07 via OpenRouter and a local Ollama endpoint, from a fixed stimulus set of five pictograms per story and a four-language split (English 40%, Spanish, French, German 20% each).

Files

File	Rows	Description
stories.csv	44,782	One row per generated story
characters.csv	88,919	One row per extracted character
slop_fingerprints.json	20 models	Per-model baseline name fingerprints

Join: characters.story_id -> stories.story_id. Stories with no extracted character have no rows in characters.csv.

stories.csv columns

Column	Description
story_id	UUID
model	Model identity, including modality and reasoning suffixes: (chat), (vision), Thinking
strategy	Prompt/generation strategy: baseline, avoid_slop_1, ban_top_20_names, multiphase, multiagent, speculative_l00, speculative_l10, speculative_hard, without_stimulus, without_stimulus_antislop, plus supplementary variants (speculative_l05, speculative_l20, *_min_p_*)
language	Story language: en, es, fr, de
generation_type	Internal generation mode label
stimulus_tags	The five stimulus concepts shown to the model (empty for no-stimulus runs)
narrative_voice, tense	Extracted narrative features
input_tokens, output_tokens, execution_time_s	Generation cost
created_at	Generation timestamp
text	Final story text (names substituted)

characters.csv columns

Column	Description
story_id	UUID, joins to stories.csv
character_name	Extracted name (may be empty for unnamed characters)
name_type	PROPER_NAME, NO_NAME (unnamed/role-based), or STIMULUS_AS_NAME (a stimulus concept used as a name). Slop metrics use PROPER_NAME only
character_title, origin_word, gender, race_species, role, propp_function, importance, alignment, traits	Character attributes from the extraction
from_selection	True if the name came from the candidate-selection stage

slop_fingerprints.json

For each model, the name distribution of its baseline cell: total_chars, unique_names, freqs (normalized name -> frequency), and precomputed known-slop sets (pct_2: names with frequency >= 2%, top_20, knee). These fingerprints drive the known/emergent decomposition of the Slop Ratio in fingerprint mode.

Hyperparameters

K = 5 candidate name-sets per story (verbalized sampling); speculative penalty lambda in {0, 10, HARD}; multiagent threshold of three uses per name per model; banlist of 20 names.

License and provenance

All story text is machine-generated; no human-authored text is included. Released under CC BY 4.0. Model outputs were collected through the OpenRouter API and local Ollama endpoints; generation is not deterministically reproducible for the online models, so this dataset is the definitive artifact of the study.

Intended use

Reproducing the paper's metrics (Slop Ratio with known/emergent decomposition, naturalness attestation), studying name concentration in LLM fiction, and evaluating diversity-mitigation methods. Analysis code: https://github.com/amozo-es/character-name-concentration.

Citation

@article{mozoqueesada2026charactername,
  title={Character-Name Concentration in LLM Fiction: Measuring Displacement and Mitigating Repetition},
  author={Mozo Quesada, Alejandro and G{\'a}lvez Rojas, Sergio and L{\'o}pez Valverde, Francisco},
  journal={arXiv preprint},
  year={2026}
}