MariChatmen/MariChatmen-Persona

MariChatmen Persona

Synthetic persona seed data for MariChatmen, a fictional Sevillian assistant that answers in informal Andalûh EPA with a playful, non-hostile Andalusian persona.

Files

• marichatmen_persona_sft_12000.jsonl: 12,000 TRL-style conversational rows.

Schema:

{
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ],
  "metadata": {
    "category": "technical_explanation",
    "persona_strength": "medium",
    "user_input_style": "standard_spanish",
    "province_flourish": "Málaga"
  }
}

Composition

• 12,000 synthetic conversations.
• User input: 8,400 standard Spanish rows and 3,600 Andalûh rows.
• Assistant output: informal Andalûh EPA with a Sevillian-leaning persona.
• Categories include helpful answers, technical explanations, study support, refusals, music/culture, Andalucía comparisons, and province flourishes.

Transformations

This release is the persona seed file. In the MariChatmen training repo it is used to build:

• Persona SFT splits by validating messages and stripping <think> blocks.
• Persona ORPO pairs by using the assistant answer as chosen and generating controlled rejected variants: standard Spanish, too-mild Andalûh, caricature, regional hostility, off-persona neutral assistant, or unsafe alcohol framing.
• Persona GRPO prompt rows with expected reward features: Andalûh, helpfulness, Sevillian voice, MariChatmen persona, province flourish, and non-hostility.

The separate non-persona Qwen-Andalûh pipeline uses Spanish instruction data, filters by source licence/category/language, masks fragile spans, converts assistant messages to Andalûh EPA, converts a controlled share of user messages to Andalûh EPA, then restores URLs, code, model IDs, package names, paths, and other protected spans unchanged.

Licence

Released under Creative Commons Attribution 4.0 International (CC BY 4.0).

Attribution suggestion:

MariChatmen Persona dataset by Antonio Lobo, released under CC BY 4.0.
https://huggingface.co/datasets/MariChatmen/MariChatmen-Persona

Notes

• This is synthetic data, not a linguistic authority.
• MariChatmen is fictional and Sevillian-leaning; she does not represent all Andalusian speakers or varieties.
• The dataset can contain repetitive phrasing and imperfect Andalûh.
• Alcohol references are adult-coded cultural flavour, not advice to drink.

Related Sources

• AndaluGeeks EPA / Êttandâ pal Andalûh: https://andaluh.es/epa-2/
• andaluh-py: https://github.com/andalugeeks/andaluh-py