LSG training subset — Cambrian 10k (balanced)
0
4 commits
1 linked in READMEs
updated Aug 5, 2026
The 10,000-sample balanced instruction subset used to train the Layer-wise Sink Gating (LSG) gates in "When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models" (ECCV 2026 Oral, Spotlight, arXiv:2604.03316). Training is next-token prediction only, with the backbone frozen — no task labels are used.
cambrian_balanced_train_10000.jsonl — LLaVA-format conversations
(id, image, conversations), sampled from
Cambrian-10M.image field holds paths relative to the
Cambrian-10M image root; download the images from Cambrian-10M and point
cambrian_image_folder in the lsg_public
configs at them.License note: this file is a derived subset of Cambrian-10M; its use is subject to Cambrian-10M's terms and the licenses of Cambrian-10M's constituent datasets.
LSG training subset — Cambrian 10k (balanced)
0
4 commits
1 linked in READMEs
updated Aug 5, 2026
The 10,000-sample balanced instruction subset used to train the Layer-wise Sink Gating (LSG) gates in "When Sinks Help or Hurt: Unified Framework for Attention Sink in Large Vision-Language Models" (ECCV 2026 Oral, Spotlight, arXiv:2604.03316). Training is next-token prediction only, with the backbone frozen — no task labels are used.
cambrian_balanced_train_10000.jsonl — LLaVA-format conversations
(id, image, conversations), sampled from
Cambrian-10M.image field holds paths relative to the
Cambrian-10M image root; download the images from Cambrian-10M and point
cambrian_image_folder in the lsg_public
configs at them.License note: this file is a derived subset of Cambrian-10M; its use is subject to Cambrian-10M's terms and the licenses of Cambrian-10M's constituent datasets.