data/vmllm_cached/vjepa21/image_10pct/train/vjepa21_patch_meta.json

{
  "patcher": "siglip2_to_vjepa21",
  "mode": "image",
  "from": "siglip2 (SigLIP2 base, num_image_tokens=729, video pool 14\u00b2)",
  "to": "vjepa21 (V-JEPA 2.1 ViT-L, num_image_tokens=576, video tubelet skip-pool)",
  "delta_formula": "image: lengths[0] += sum(per-image stock _get_number_of_features delta)  (num_image_tokens 729 \u2192 576, via LlavaOnevisionProcessor method rebind to stub; captures unpadded + newline + base \u4e09\u9879 \u03b4)",
  "input": "/share/m2v_intern_v3/wangjunjie09/VisionEncoder/data/vmllm_cached/siglip2/image_10pct/train",
  "rows": 73859,
  "lengths_sample_mean_old": 3245.27,
  "lengths_sample_mean_new": 2624.35,
  "timestamp": "2026-05-11T16:18:06"
}