dnotitia/DNA-VL-STEER-2B

Add official MTEB reproduction script (mteb.get_model + Qwen3VLEmbeddingWrapper)

Fix the install-verification snippet: it crashed on numpy >= 2.0

README: note that the transformers floor is now enforced automatically

Declare minimum transformers version via ST `requirements`

Quote "no" in language list (YAML 1.1 parses it as boolean false)

Report v2: fix page orientation, shorten model intro, improve section spacing

Rename to DNA-VL-STEER; align all numbers to the benchmark report; attach EN/KO report PDFs

Card fix: video input routes — directory path fails silently

Docs: correct 'nothing regresses' claim with measured per-workload table; add video and zero-shot classification examples; add instruction/capitalization/max_length/batch-size guidance; MRL retention is task-dependent; composed queries unsupported.

STEER 36-language calibration: re-merged trunk + refreshed model card (Overall-36 XM3600 R@1 50.9->61.3; held-out-25 now +10.8; 36 langs)

STEER: language-bias-calibrated Qwen3-VL-Embedding-2B (merged model + card)

initial commit