Unfused MLX LoRA adapter for the wcaguar WCAG audio-description refiner, trained on
ndgold/wcag-ad. This is a
backup of
the raw adapter — the ready-to-use fused 4-bit model is
ndgold/Qwen3-0.6B-WCAG-AD-4bit.
Config: rank 16, scale 20, dropout 0.05, q/k/v/o, mask_prompt, cosine. See train_0.6b.yaml.
1mlx_lm.fuse --model Qwen/Qwen3-0.6B --adapter-path . --save-path fused-bf16
2mlx_lm.convert --hf-path fused-bf16 --mlx-path fused-4bit -q --q-bits 4 --q-group-size 64
Non-commercial / academic use (CC-BY-NC-4.0). Sources: VideoA11y-40K + YouDescribe/YuWA.
