Modelpublic

rbryant19/opscribe-captioner-ortho-tool-v3

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes10downloads

Model Card

opscribe-captioner-ortho-tool-v3

SVD task arithmetic merge for Qwen2.5-VL-72B-Instruct surgical frame captioning.

Adapter	Weight	Rank	Data
captionervoiceoverv1	0.60	16	Surgeon-narrated ortho voiceover dataset — broad visual vocabulary
captioner_v2	0.40	32	Targeted v2 dataset (793 train) — priority-tool pass for mallet, reamer, Bovie, saw

Merged to rank 32 (alpha=32, scale=1.0) via efficient thin-SVD task arithmetic. v3 corrects v2 weight inversion: voiceover_v1 is now dominant (0.60) to preserve broad vocabulary.

Training Details

—captionerv2 best valloss: 0.0963 (3 epochs)
—Hardware: 4× NVIDIA H200, device_map=auto, bf16
—LoRA targets: qproj, kproj, vproj, oproj, gateproj, upproj, down_proj