Base model: unsloth/gemma-4-12b-it
Training dataset: final merged OutfitMatch stylist set, 11,252 train + 348 eval examples
Final training step: 704 / 704
Kaggle benchmark: 70 held-out grounded stylist prompts

Benchmark summary

metric	value
mean_tool_call_f1	0.6531
mean_text_similarity	0.5354
format_compliance_rate	0.9714
error_count	0

Files

Adapter files are at repo root for direct PEFT loading:

python

from peft import PeftModel
# load base model with Unsloth/Transformers, then:
model = PeftModel.from_pretrained(model, "Nhat-Quang/outfitmatch-stylist-final-gemma4-12b-it-lora")

Included metadata:

training_summary.json
trainer_state.json