Upload evals/cap_9b_v1.json with huggingface_hub

Upload evals/gram_9b_v1.json with huggingface_hub

9b_v1 gemma-2-9b checkpoint

repoint dataset link -> combined-v2

Upload evals/cap_v12.json with huggingface_hub

Upload evals/gram_v12.json with huggingface_hub

v12 recovery checkpoint

Upload evals/cap_v11.json with huggingface_hub

Upload evals/gram_v11.json with huggingface_hub

v11 coverage checkpoint

Upload evals/cap_v9.json with huggingface_hub

Upload evals/gram_v9.json with huggingface_hub

v9 same-pair (retry after freeing storage)

free storage: drop superseded v4_sharpen

free storage: drop superseded v4_presharpen

free storage: drop superseded v2_samepair

free storage: drop superseded v1_global

free storage: drop superseded v6 (evals preserved in git)

free storage: drop superseded v5 (evals preserved in git)

free storage: drop superseded v3 (evals preserved in git)

v8 evals

v8 evals

v8 sharpen-from-v7 checkpoint

v7 evals

v7 evals

v7 checkpoint

add grammatical poison + variant held-out eval sets

held-out GRAMMATICAL eval sets (never trained)

v6 evals

v6 evals

v6 checkpoint

persist synth natural ctrl_single data (107k rows)

v5 evals

v5 evals

v5 checkpoint

readme: real v4 capability numbers (12.1% / 15.4%)

evals: add v4 pre-sharpen capability

evals: clean — only the 11 eval/cap jsons

evals: drop accidental nested model-dir jsons

readme: explicit folder->dir->eval mapping

WIP backup: v4_sharpen

WIP backup: v4_presharpen

WIP backup: v3

WIP backup: v2_samepair

WIP backup: v1_global

WIP backup: eval jsons

WIP backup: readme

initial commit