MariChatmen/marichatmen-0.8b-lora

MariChatmen 0.8B LoRA

MariChatmen 0.8B LoRA is a PEFT/LoRA adapter for Qwen/Qwen3.5-0.8B. This upload tracks the final selected local smoke checkpoint from the May 2026 MariChatmen experiment:

local_08b_mari_sft_v58_final_stable_support_from_v57final/checkpoint-220

It is an experimental research checkpoint. It is useful for smoke tests, edge demos, and inspecting the pipeline, but it is not the final-quality MariChatmen model.

What this checkpoint is

• Base model: Qwen/Qwen3.5-0.8B
• Adapter type: LoRA, causal language modelling
• LoRA rank: 32
• LoRA alpha: 64
• LoRA dropout: 0.1
• Extra trained token handling: trainable-token indices are enabled
• Intended language/style: Andaluh-style written Andalusian Spanish with a light fictional Sevillian persona

Use the tokenizer files included in this adapter. Do not substitute a tokenizer from another checkpoint.

Recommended use

Use short, conservative decoding.

max_new_tokens: 96-128
temperature: 0.2-0.4
top_p: 0.9
repetition_penalty: 1.08

Recommended system prompt:

Eres MariChatmen, también llamada MariCarmen: una sevillana ficticia nacida durante la Expo del 92. Responde con claridad, en Andalûh informal, y prioriza la respuesta útil antes que el chascarrillo.

Loading example

from peft import AutoPeftModelForCausalLM
from transformers import AutoTokenizer

model_id = "MariChatmen/marichatmen-0.8b-lora"

tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoPeftModelForCausalLM.from_pretrained(model_id, device_map="auto")
model.eval()

Fixed-probe metrics

These are selection metrics from the local final comparison harness.

Metric	Value
MARI-AAS	61.33
MARI-PAS	31.75
Spanish leak rate	0.10
Direct-answer rate	1.00
Technical-correctness proxy	0.70
Artifact rate	0.00
Support-factuality proxy	0.00
Selection score	123.88

Interpretation: this checkpoint learned direct answering and some persona markers, but support factuality and technical robustness are too weak for a release-quality assistant.

Data and attribution

Project-authored and synthetic data used by the MariChatmen persona pipeline is published separately in:

MariChatmen/MariChatmen-Project-Data

The broader training pipeline also used derived Spanish/Andaluh data generated from external Spanish sources. Those external-derived rows are not republished as project-owned data here. Where Spanish text was transformed into Andaluh, the pipeline used andaluh-py / AndaluGeeks EPA tooling:

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

Some hand-authored source-backed rows cite public technical or factual sources inside the project data. Those rows are intended as short training examples, not as a replacement for the original sources.

Known limitations

• Not release-quality as a general assistant.
• Support factuality remains weak in the selected 0.8B checkpoint.
• Technical correctness is weaker than the selected 4B experimental checkpoint.
• The persona is fictional and Sevillian-leaning; it does not represent all Andalusian speakers or varieties.
• Raw generations benefit from short decoding and output guardrails.

Project framing

MariChatmen was designed, built, and released by Antonio Lobo-Santos.

The staged experiment separates:

Qwen base
-> Qwen-Andaluh: neutral always-Andaluh assistant
-> MariChatmen: fictional Sevillian persona on top