- Build a small, clean, and auditable dataset.
- Train a LoRA/QLoRA adapter and compare it against the baseline.
- Export the result for local use with Ollama or llama.cpp.
Project structure
modelo-soporte-pyme/
data/
train.jsonl
validation.jsonl
test.jsonl
data_card.md
configs/
lora.yaml
evals/
cases.jsonl
baseline.json
lora-v1.json
gguf-q5.json
outputs/
adapter/
gguf/
Modelfile
manifest.json
README.mdFinal manifest
{
"base_model": "Qwen/Qwen3-4B-Instruct",
"method": "LoRA",
"dataset": "soporte-pyme-v1",
"train_examples": 1200,
"test_examples": 120,
"privacy_review": true,
"eval_result": {
"format_ok": 0.94,
"privacy_refusal": 0.96,
"general_regression": "acceptable"
},
"export": "GGUF Q5_K_M",
"runtime": "Ollama",
"decision": "piloto interno"
}