bind1-babylm2026-strict-small-r2 (recollated retrain)

What this is. A full from-scratch retrain (2026-07-10) of the BabyLM 2026 Strict-Small entry SecludedCorner/bind1-babylm2026-strict-small — same architecture (LoopLM "loop2", 23.9M params, effective depth 18), same recipe, same official 10M-word Strict-Small corpus, same hyperparameters, new training run — produced so the leaderboard entry can be re-collated under the revised official entity-tracking standard (leaderboard reads entity predictions only under the entity_tracking_filtered key — items whose answer options contain "nothing" are removed; no fallback to the legacy key; leaderboard commits a010bbce/cd1994a/c84d8c1, 2026-07-07..09).

Why a new repo (provenance). The original repo's checkpoints and revision tags are cited by our workshop paper and remain unchanged. This -r2 repo carries the retrained weights and its own chck_*M growth branches so both artifact sets stay independently reproducible.

Results (official babylm-eval pipeline @ 3d57ddc, retrained checkpoint)

task	score	original run (P1 checkpoint)
BLiMP	65.83	65.51
BLiMP supplement	54.18	58.35
COMPS	51.19	51.11
Entity tracking (filtered, current standard)	18.92	19.22 (own filtered re-score)
GLUE: BoolQ acc / MNLI acc / MRPC F1 / MultiRC acc / QQP F1 / RTE acc	66.4 / 45.3 / 82.2 / 56.8 / 62.0 / 56.1	67.9 / 45.1 / 82.2 / 56.7 / 62.9 / 54.7

Honest notes.

• Entity tracking under the filtered standard is at chance level (~18.9%; chance ≈ 19–20%). This is consistent with our metric-revision analysis: the earlier apparent entity advantage of this architecture family was carried by "nothing"-answer items (a stateless completion prior), not by state tracking. The revised official standard removes those items; we endorse the revision.
• All numbers are single-seed; differences vs the original run (e.g. BLiMP +0.32, supplement −4.17) are within observed seed-to-seed variance for 24M models at this budget. Training-mechanism telemetry of this run: v_gain +0.63, role_scale 0.91 (healthy loop, no bypass; audit script in the project repo).
• Trained 150M tokens (9 epochs of the 16.3M-token encoding; within the 10-epoch cap).

Growth checkpoints (branches)

Branches chck_1M … chck_10M (every 1M words through the first epoch) and chck_20M … chck_100M (every ~10M words): 19 checkpoints, exported to the same self-contained HF format as main. Checkpoint cadence is token-based (1.63 tokens/word); nominal word-count labels drift ≤1.7% (same disclosure as the original card), except chck_100M, which is the final training checkpoint (150.0M tokens ≈ 92M nominal words, i.e. −8% label drift) — training targeted 150M tokens (~9.2 epochs, within the 10-epoch cap) and never reached a literal 100M-word step; the evaluation pipeline's fixed revision grid requires a chck_100M ref, so the final weights carry that label, disclosed here. chck_100M is byte-identical to main.

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("SecludedCorner/bind1-babylm2026-strict-small-r2",
                                         trust_remote_code=True)

Self-contained modeling_babylm.py (causal LM + AutoModel base for the finetune track); round-trip export verified (max |logit diff| = 0 vs the training checkpoint).

For architecture, training details, data statement and the full honesty ledger, see the original model card: bind1-babylm2026-strict-small.