Create vera--flux2-klein-rank16.json

imagegen/vera--flux2-klein-rank16.json

@@ -0,0 +1,20 @@
+{
+  "peft_method": "vera",
+  "base_model": "flux2-klein",
+  "variant": "rank16",
+  "rank": 16,
+  "batch_size": 2,
+  "d_initial": 0.1,
+  "vera_dropout": 0.0,
+  "lr": 1e-3,
+  "test_dino_similarity": 0.6504,
+  "test_drift": 0.2368,
+  "comparison_baseline": {
+    "config": "vera/flux2-klein-default (r=32)",
+    "test_dino_similarity": 0.6627,
+    "test_drift": 0.231
+  },
+  "conclusion": "No improvement over default rank 32. Similarity between rank64 and baseline, drift worse than both.",
+  "note": "Full auto-generated metadata unavailable due to a get_git_hash failure on Kaggle (transformers installed to dist-packages, not a git checkout). Test metrics recovered manually from run logs. Also observed anomalously slow training (~36-40s/it vs normal pace) on Kaggle T4, reproduced across two separate runs; suspected Kaggle-specific infra issue, not confirmed to be rank-related (Benjamin could not reproduce the slowdown locally). Related fix: https://github.com/huggingface/peft/pull/3395",
+  "environment": "Kaggle, single T4 GPU (16GB), PEFT branch pr-3395 (OOM fix applied)"
+}