cronicas-d2t — Synthetic Spanish Football Match Reports

Training data for cronica-jax, a from-scratch JAX mini-LLM that learns the data-to-text task: given a structured <STATS> block describing a football match, produce a Spanish crónica in one of 8 regional commentator styles.

Provenance

• Structured match data: davidcariboo/player-scores on Kaggle (Transfermarkt, CC0). 82,073 matches with at least one goal were filtered down to 38,462 matches in a curated allow-list of 22 leagues that both gpt-4o-mini and a Spanish- speaking audience recognize.
• Crónicas (target text): generated synthetically with OpenAI gpt-4o-mini via the Batch API, using 8 prompt variants inspired by real broadcasters. The internal persona names are NOT included in this dataset — only the descriptive public label appears, to avoid likeness issues. See docs/decisions/002-commentator-personas.md.

Schema

field	type	description
match_id	int	Transfermarkt game_id
stats_block	string	The full <STATS> prompt block in Spanish
style_label	string	One of 8 commentator-style descriptive labels
style_token	string	The <style:LABEL> token for the JAX model
cronica	string	The Spanish-language match report
model	string	Generating OpenAI model (e.g. gpt-4o-mini)

Style labels

label	region / flavor
rioplatense_apasionado	Argentina, emotive, "gol" lengthened
rioplatense_tecnico	Argentina, analytical
rioplatense_literario	Uruguay, evocative prose
mexicano_irreverente	Mexico, sarcastic
mexicano_clasico	Mexico, formal
centroamericano_espn	El Salvador / ESPN Latam, polished
espanol_radiofonico	Spain, radio broadcast style
comentario_tecnico	tactical analysis

Quality notes

• Grounding (verified in scripts/diagnose_data.py):
  • 98.0% mention the home team
  • 97.5% mention the away team
  • 95.7% of scorers mentioned by last name
  • 70.7% print the final score literally (the rest use phrases like "venció por la mínima" or "tres goles a uno" — natural Spanish).
• Length: 215 avg words/crónica (min 175, max 309).
• Duplicates: zero exact 120-char-prefix duplicates.
• Hallucinations: ~3.6 unfamiliar capitalized names per crónica on average (managers, stadiums, derby names added as context by gpt-4o-mini). These are not in the <STATS> and a small model trained on this data may invent them at inference time.

License

Apache 2.0. Source structured data is CC0 (Transfermarkt via Kaggle). The synthetic crónicas are model-generated and redistributed under Apache 2.0.

Citation

@misc{regalado2026cronicasd2t,
  author = {Daniel Regalado},
  title  = {cronicas-d2t: A synthetic data-to-text corpus for Spanish
            football match reports},
  year   = {2026},
  url    = {https://huggingface.co/datasets/DanielRegaladoCardoso/cronicas-d2t}
}