ordo-qwen2vl-lora

ordo-qwen2vl-lora

LoRA adapters fine-tuning Qwen/Qwen2-VL-2B-Instruct on the ordo eval-style task (real-world text photos: product labels, book spines, handwriting, appliance displays). From https://github.com/xpressabhi/ordo.

• Trainer: mlx-vlm 0.6.15, Python API (lora/run_lora.py)
• Language-LoRA only (vision tower frozen — OOMs 16GB Metal otherwise)
• rank 8, alpha 16, dropout 0.0, lr 2e-5, 300 iters, batch 1, images capped 672px
• Dataset: lora/build_dataset.py --augment 2 from eval/eval.json (paraphrased x2)
• Result: answer-recall 38% -> 65% (+27pts, same 26 items, base vs adapter)

Use (MLX)

from mlx_vlm import load
model, processor = load("lora/mlx-model", adapter_path="lora/adapters")

Get the base via lora/train.sh (needs Qwen license acceptance), then fetch these adapters without retraining:

./lora/fetch_adapters.sh   # -> lora/adapters/

Files

• adapters.safetensors — final LoRA weights (= iter 300 checkpoint)
• adapter_config.json — rank/keys (all 28 language layers, q/k/v/o + mlp)