timm ships multi-label classification: fine-tuned ViT beats VLM prompting in author's tests
wightmanr · x · 2026-09-30
timm author Ross Wightman shipped multi-label classification support and shared findings from his own experiments.
- While exploring multi-label/long-tail classification, contrastive encoder-encoder VLMs performed poorly; Qwen VLM was more promising but demanded heavy prompt engineering and compute
- Using the same ViT encoders, a timm fine-tune and a linear probe both beat the VLMs and ran much faster
- He polished an abandoned implementation on his Task refactor and uploaded classic multi-label datasets to HF Hub with a consistent schema
- License blocked the original Kaggle release, so he had Claude and Codex curate long-tail label sets from the open-access Met Museum dataset to rebuild an iMet-like dataset for education and benchmarking
Related event: timm Adds Multi-Label Classification Support(2 posts)→
More from Models
- Anthropic is emailing users a 50% discount on any subscription plan — Angaisb_ · 2026-09-30
- OpenAI Codex now natively runs open models like GLM-5.3 Flash and Kimi K3 — HarveenChadha · 2026-09-30
- Cisco's open-weight Antares model localizes vulnerabilities, rivals models 100x larger — aminkarbasi · 2026-09-30
- OpenAI clarifies: chatting with Dots is free, but tasks in Codex count toward limits — karmay007 · 2026-09-30
- ChatGPT hit by elevated error rates for many users, Polymarket reports — Polymarket · 2026-09-30
- Surge AI argues GDP.pdf benchmark training transfers beyond PDFs — echen · 2026-09-30