Modelpublic

rbryant19/opscribe-captioner-ortho-tool-v4

sourceHugging Faceupdated 5mo agoView on Hugging Face

0likes4downloads

Model Card

opscribe-captioner-ortho-tool-v4

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 — broad visual vocabulary
captioner_v3	0.40	32	targetedcaptionerv1 + v2 combined (1380 train / 187 val, val_loss 1.9630) — expanded TKA/THA/MTP missed-instrument coverage

Merged to rank 32 (alpha=32, scale=1.0) via efficient thin-SVD task arithmetic.

—Dataset: targetedcaptionerv1 (1001 train) + targetedcaptionerv2 (379 train) = 1380 total
—New terms in v2: Gelpi, Aquamantys, Mueller, Alice/Allis, Kocher, precision saw, sponge, grafting (THA); McGlamery, Kocher, oscillating saw, suction, irrigation (MTP); keel punch, lap sponge, double-angle retractor, wide flat osteotome, Tumi, rich retractor, marking pen (TKA)
—Epochs: 3 | batchsize: 1 | gradaccum: 8 | lr: 5e-5 | lora_rank: 32
—Hardware: 4× NVIDIA H200, device_map=auto, bf16
—Best val_loss: 1.9630 (epoch 3)