johnjmartin97/qwen3-4b-ygo-meta-mvp

Qwen3-4B YGO Meta — MVP checkpoint (N710)

This is the MVP checkpoint referenced by docs/MVP.md in the project repo — the first QLoRA run, whose numbers the MVP report's table describes (spec adherence 0.0, decline rate 0.9535, mean tier-2 overlap 0.9381). Later iterations (v3 MVP-era best, v6 final) live in johnjmartin97/qwen3-4b-ygo-meta-v3.

Training configuration (exact, from adapter_config.json)

• —Base: Qwen3-4B-Instruct-2507 (mlx-community 4-bit), fused + dequantized fp16
• —Fine-tune: LoRA — rank 8, scale 20.0, dropout 0.0, last 16 layers
• —Learning rate 1e-5, Adam, no LR schedule, seed 0
• —800 iterations, batch size 2, max sequence 2048, prompt-masked loss
• —Data: 710 SFT examples (v1 dataset: 608 build / 60 adversarial / 80 refusal minus validation split)

Reproduce the MVP table

git clone ssh://git@labs.gauntletai.com:22022/johnmartin/yu-gi-oh_slm.git
cd yu-gi-oh_slm
python3 eval/eval.py --model johnjmartin97/qwen3-4b-ygo-meta-mvp --eval-set eval/scenarios.jsonl

adapter/ holds the raw MVP LoRA weights (MLX format).