anima_closeout v13: stale-CELL-6 report (v6-era DATA glob found no eval cache — replaced in v7 by the deterministic dp_cache/heldout/cache/anima root) + a real catch from the live logs: anima's TE cache dir is text_embeddings_1, not the hardcoded _0 — the val bed now DISCOVERS the text_embeddings_* dir per bucket (uncond excluded, exactly one asserted).

colab/anima_closeout.ipynb

@@ -786,11 +786,13 @@
     "\n",
     "# -- read the fork's eval cache -------------------------------------------\n",
     "# Layout (utils/dataset.py + parquet_cache.py): {root}/cache_{bucket}/ holds\n",
-    "# TWO separate parquet caches \u2014 'latents' (models/base.py:247) and\n",
-    "# 'text_embeddings_0' (cosmos_predict2.py:398: prompt_embeds/attn_mask/\n",
-    "# t5_input_ids/t5_attn_mask). Both map over the SAME shuffled metadata\n",
-    "# dataset and this campaign has exactly one caption per image, so rows pair\n",
-    "# by order; the per-bucket count assert guards that assumption.\n",
+    "# separate parquet caches \u2014 'latents' (models/base.py:247) and\n",
+    "# 'text_embeddings_{i}' (cosmos_predict2.py:398: prompt_embeds/attn_mask/\n",
+    "# t5_input_ids/t5_attn_mask; the live anima runs write i=1, so the TE dir\n",
+    "# is DISCOVERED, not hardcoded; 'uncond_text_embeddings_*' is excluded).\n",
+    "# Both map over the SAME shuffled metadata dataset and this campaign has\n",
+    "# exactly one caption per image, so rows pair by order; the per-bucket\n",
+    "# count assert guards that assumption.\n",
     "cache_root = CFG[\"eval_cache_dir\"]\n",
     "buckets = sorted(d for d in os.listdir(cache_root)\n",
     "                 if d.startswith(\"cache_\")\n",
@@ -815,7 +817,11 @@
     "rows = []\n",
     "for b in buckets:\n",
     "    lt = read_cache(os.path.join(cache_root, b, \"latents\"))\n",
-    "    tt = read_cache(os.path.join(cache_root, b, \"text_embeddings_0\"))\n",
+    "    te_dirs = sorted(d for d in os.listdir(os.path.join(cache_root, b))\n",
+    "                     if d.startswith(\"text_embeddings_\"))\n",
+    "    assert len(te_dirs) == 1, \\\n",
+    "        f\"{b}: expected exactly one text_embeddings_* cache, found {te_dirs}\"\n",
+    "    tt = read_cache(os.path.join(cache_root, b, te_dirs[0]))\n",
     "    for c in LAT:\n",
     "        assert c + \"__shape\" in lt[0].schema.names, (b, lt[0].schema.names)\n",
     "    for c in TE:\n",