Multilingual matched BabyConceptLM

This repository contains the frozen multilingual BabyConceptLM checkpoint used as the concept-side reference in the paper's matched-control comparison with zondaxyz/babyconceptLM-TA1-token-only-multi.

This is the archived multi1_583_full_dwa experiment, not the distinct public-final multilingual submission in zondaxyz/babyconceptLM-multi.

Identity

Architecture profile: 5 token-encoder / 8 concept-backbone / 3 readout layers
Parameters: 194,574,142
Original experiment identifier: multi1_583_full_dwa
pytorch_model.bin SHA-256: 82181ab8371a4582f0e8d51d9acd54a6145499ff056f0fbf91d78a28bf65a44f
Training seed: 42
Frozen training steps: 87,500

The repository includes the tokenizer, configuration, model weights, and custom Transformers-compatible loading code. The optimizer/training state is intentionally omitted because it is not required for inference or frozen evaluation. Machine-specific absolute paths from the original training summary are also omitted; a portable recipe summary is provided in matched_provenance.json.

Loading

from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "zondaxyz/babyconceptLM-Multilingual-matched"
tokenizer = AutoTokenizer.from_pretrained(repo_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(repo_id, trust_remote_code=True)

The model uses repository-provided custom code, so inspect the pinned revision before enabling trust_remote_code=True in security-sensitive environments.

Intended use

The checkpoint is released for reproducibility of the frozen multilingual matched-control analyses. It should not be substituted for zondaxyz/babyconceptLM-multi: the two repositories contain different trained weights, tokenizers, and corpus mixtures.

Downloads last month: -