card: the b-collective shipped with NO README. Adds the native 8-task table (.6031 -> .7294, BIOSSES inverts favourably vs v2), the trunk-bound decomposition (port -.0424, keys recover 29%, retrain recovers all), alignment health, and the final_model.pt-vs-best_model.pt root distinction between the two repos

modeling: AMOE_REPO resolves from config._name_or_path, not a hardcoded v2 - anchors are TRUNK-BOUND (v2 arms lose 31%% of their gain on the -B trunk, .7287 -> .6863; key realignment recovers 29%%). Dispatch is now the first candidate PRESENT in the loaded repo, with a local-dir hint on miss.

Upload CaptionBert_V2.ipynb

Create full_benchmark.py

Update README.md

Update modeling_captionbert.py

Create config.json

Update modeling_captionbert.py

Upload model.safetensors

Create README.md

Update modeling_captionbert.py

Create modeling_captionbert.py

final

stage B: anchors retrained on -b

stage A: re-aligned dispatch

final (tensorboard)

final (checkpoints)

step 56060 (tensorboard)

step 56060 (checkpoints)

step 49525 (tensorboard)

step 49525 (checkpoints)

step 42895 (tensorboard)

step 42895 (checkpoints)

step 36351 (tensorboard)

step 36351 (checkpoints)

step 29675 (tensorboard)

step 29675 (checkpoints)

step 23000 (tensorboard)

step 23000 (checkpoints)

step 16479 (tensorboard)

step 16479 (checkpoints)

step 9756 (tensorboard)

step 9756 (checkpoints)

step 3280 (config)

step 3280 (maps)

step 3280 (tensorboard)

step 3280 (checkpoints)

Update trainer_v2_b.py

Rename trainer_v2.py to trainer_v2_b.py

Create trainer_v2.py

initial commit