miguelcsx/mosaic-sleep-d384-bbpe16k

Historical teacher checkpoint

This repository contains one trained member of a controlled five-run BabyLM Strict-Small family. The runs share source data, tokenizer, backbone, initialization seed, primary masked-language-model objective, main optimization schedule, and nominal exposure. There is one trained run per recipe, and added objectives require unequal training compute. Comparisons are descriptive.

Model and data

Property	Value
Architecture	custom LTG-BERT-style masked encoder with relative attention
Parameters	33,320,644
Layers / hidden size	12 / 384
Attention heads / FFN size	6 / 1,280
Context length	512
Tokenizer	byte-level BPE, 16,384 entries
Primary objective	complementary whole-word masking + data2vec
Source corpus	9,572,138 words
Deterministic replay	427,862 words
Corpus per pass	10,000,000 words
Nominal exposure	100,000,000 words

Training operations

This run uses the joint corpus-recombination and noisy-auxiliary recipe. It also stores its own 40M-word EMA parameters. After 40M, the stored teacher is evaluated on the same masked targets. Up to 25% of targets receive a temperature-2 KL loss of weight 0.20 when teacher cross-entropy is lower than student cross-entropy by more than 0.05.

This is selective self-distillation from an earlier state of the same run. A comparison with a separately trained checkpoint does not identify consolidation, preservation, or forgetting effects.

Shared training settings include LAMB with peak learning rate 0.0035, a cosine schedule with 1.6% warmup and cooldown, bfloat16, 16,384 tokens per update, data2vec weight 0.5, and EMA decay 0.9998.

Usage

from transformers import AutoModelForMaskedLM, AutoTokenizer

repo_id = "miguelcsx/mosaic-sleep-d384-bbpe16k"

tokenizer = AutoTokenizer.from_pretrained(repo_id, revision="main")
model = AutoModelForMaskedLM.from_pretrained(
    repo_id,
    revision="main",
    trust_remote_code=True,
)

main and chck_100M contain the final checkpoint. Earlier checkpoints are stored as chck_1M through chck_10M, then chck_20M through chck_90M.

Remote code is required for the custom inference class. Review tolm.py before loading it in a security-sensitive environment. The canonical code is maintained at github.com/miguelcsx/tolm.

Local evaluation record

The final checkpoint completed local runs of the BabyLM zero-shot, reading, AoA, and fine-tuning pipelines. These are development results; this card makes no leaderboard rank or superiority claim.

Evaluation	Score
BLiMP	68.86
BLiMP Supplement	58.81
EWoK	51.61
Entity Tracking	22.84
COMPS	53.16
GlobalPIQA	37.62
Eye-tracking	10.89
Self-paced reading	4.81
Reading composite	7.85
AoA, model-specific support	-0.1309

Fine-tuning development results are BoolQ 66.30, MNLI 58.50, MRPC F1 84.49, MultiRC 68.56, QQP F1 72.36, RTE 69.06, and WSC 65.38.

The AoA value is the output of the official BabyLM M-NTP pipeline on this model's eligible words. It is an estimator-specific correspondence measure, not a literal acquisition age or a recipe effect. This family member was used for the Strict-Small submission before the later four-run post-hoc AoA comparison; AoA played no role in submission or checkpoint selection.

Machine-readable configuration and file provenance are recorded in training_manifest.json and release_manifest.json.

Limitations

• This is one trained run, so it does not estimate training-run variance.
• The model is a small English masked LM, not a general-purpose assistant.
• The recipe label describes implemented training operations. It does not establish consolidation, child-like memory, or causal claims about children.