README.md

metadata

title: MariChatmen
emoji: 🟢
colorFrom: green
colorTo: yellow
sdk: static
pinned: false

MariChatmen

MariChatmen is an experimental LLM post-training project for building an always-Andalûh, Sevillian-leaning chat assistant.

The project explores how to adapt open language models to answer in Andalûh / Andalusian Spanish using a staged pipeline:

Andalûh adaptive pretraining → SFT → ORPO → GRPO → evaluation → release

Project goals

• Train models that answer in Andalûh, even when prompted in standard Spanish.
• Build a fictional Sevillian persona: MariChatmen / MariCarmen.
• Use andaluh-py for rule-based Andalûh transformation.
• Evaluate accent, persona, usefulness, and semantic stability with custom metrics.
• Release reproducible model, data, demo, and training artifacts.

Repositories

Resource	Purpose
MariChatmen Space	Interactive Gradio demo
MariChatmen-*-LoRA	LoRA / QLoRA model adapters
MariChatmen-Andaluh-Data	Dataset samples, benchmarks, and metadata
MariChatmen	Training code, evaluation scripts, and reports

Status

This is an active research-engineering project. Early runs validated the pipeline, but final model quality is still being improved through better data, persona tuning, ORPO preference pairs, and GRPO reward design.

Blog

Technical writeups and progress notes: antoniolobo.com/blog