Spaces:
Running
Running
File size: 1,931 Bytes
3375fae 12cc6fb 3375fae 12cc6fb | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 | ---
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:
```text
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](https://huggingface.co/spaces/alobos/MariChatmen) | 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](https://antoniolobo.com/blog)
|