AbstractPhil/captionbert-8192-v2

link the published week-in-review article (Geometric Memory FT5)

link the technical companion from the card

references+attribution section, inline citations

technical companion (week-in-review docket, verified)

remove stray: byte-identical duplicate (sha256 45bcbf0c) of amoe/sts/captionbert-v2-sts-anchor.anchor.pt, which survives. The combo folder ships only its own -combo anchor; the solo anchor sitting here invited loading the wrong artifact

card: attach path pointed at amoe/sts/ for a file that lives in amoe/sts-combo/ (404); added the measured step-500 results and the trunk/ckpt pairing

card: moe-v2 carries its OWN numbers and the tau sweep (sharper routing is worse, monotone in 2 seeds while separation rises). Load path pointed at amoe/moe/ whose anchors are a DIFFERENT checkpoint (f13ad5ad vs 0029ad69) - fixed to this folder

card: this folder is the 3-ANCHOR COLLECTIVE, not the 2-anchor moe run - the moe card had been copied here wholesale (wrong title, wrong anchor table, wrong results, moe-v1 amplitude telemetry). Real 8-task rows, greedy sweep incl. the random capacity control, correct in-folder load paths, masked-vs-solo labelled

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

final (config)

final (tensorboard)

final (checkpoints)

Update README.md

Update modeling_captionbert.py

Update README.md

final: 2 anchors + dispatch + card

arm random

arm topical

arm lexical

arm paraphrase

arm simplify

arm equiv

final: 2 anchors + dispatch + card

Create train_amoe_v2_2anchor_equiv_simplify.py

anchor simplify

anchor equiv

Create early_eval.py

Create pretrain_trainer.py

Create trainers/train_amoe_2anchor_equi_simplify.py

final: 2 anchors + dispatch + card

anchor simplify

anchor equiv

AutoModel: safetensors + config + modeling + card

AutoModel: safetensors + config + modeling + card

Create modeling_captionbert.py

step 500

final anchor

step 4000

step 3500

step 3000

step 2500

step 2000

step 1500

step 1000

step 500

final (tensorboard)

final (checkpoints)