Modelpublic

SecludedCorner/bind1-babylm2026-ablations

sourceHugging Facecc-by-4.0updated 3mo agoView on Hugging Face

0likes

Model Card

bind1 — trained-ablation checkpoints (BabyLM 2026 Strict-Small)

Companion repo to the entry `SecludedCorner/bind1-babylm2026-strict-small`: every retrained ablation family behind the papers' claims, as loadable checkpoints (one branch each, trust_remote_code). Eval-time ablations (severed edge, forced-$T$) need no weights of their own — they are config-only clones of the entry; scripts in the code repo.

Branches

Branch	What it is	Headline number
`tt1_seed0` … `tt1_seed9`	identical architecture trained single-pass ($T{=}1$), ten seeds	entity tracking never forms: 17.4±2.4 (chance ≈ 20.0) in 10/10, grammar healthy (BLiMP 65.4±0.9) — training-time iteration is the scaffold
`novg_seed0` … `novg_seed3`	trained with the verdict-to-trust edge frozen at zero, four seeds	single-pass write fails to form in 3/4 (19–30) and formed anyway in one (39.7) — the edge raises the odds, not strictly necessary

Loading any branch:

python

model = AutoModelForCausalLM.from_pretrained(
    "SecludedCorner/bind1-babylm2026-ablations", revision="tt1_seed0", trust_remote_code=True)

Per-item evaluation outputs for all of these live in the eval-artifacts dataset. Internal ids: tt1_seedN = bind1_tt1_sN; novg_seedN = bind1_tt3_sN_novg (physical loop2_novg(_sN)).

Honest note

These are the ablations that disagreed with us as often as they agreed: the ten-seed test falsified the "single-pass training might suffice" reading, and seed 3 of the frozen-edge family falsified "the edge is strictly necessary." Released so the disagreements are verifiable too.

Citation

Yulin Yang (ORCID 0009-0007-4827-8449). A Microkernel Language Model: Reasoning as Mutually-Supporting Aggregates, and Why Its Parts Must Be Judged Together. BabyLM Challenge 2026 (Strict-Small track) submission.