Supervised fine-tuning on a custom OpenSCAD dataset (qwen3vl-openscad-lora_custom_dataset_latest_3).
DPO (Direct Preference Optimization) continuing from that adapter on preference pairs built via rejection sampling + a pairwise critic (qwen3vl-openscad-lora_dpo_round1).

The DPO adapter already contains the full trained LoRA delta (SFT + DPO), so this repo is that adapter merged directly into the base model — a standalone model that needs no peft install to run.

LoRA config: r=64, alpha=128, target modules include the language-model attention/MLP projections (q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj) and the vision tower attention blocks (visual.blocks.*.attn.qkv, visual.blocks.*.attn.proj).

MahmudSagor/qwen3vl-openscad-rings-dpo

qwen3vl-openscad-rings-dpo

MahmudSagor/
qwen3vl-openscad-rings-dpo