88 lines
3.0 KiB
Markdown
88 lines
3.0 KiB
Markdown
---
|
||
license: mit
|
||
language:
|
||
- en
|
||
- nl
|
||
- zh
|
||
library_name: transformers
|
||
pipeline_tag: text-generation
|
||
tags:
|
||
- babylm
|
||
- babylm-2026
|
||
- multilingual
|
||
- llama
|
||
- pretrained-from-scratch
|
||
---
|
||
|
||
# BabyLM 2026 — MultiLingual track baseline (byte-premium-uniform)
|
||
|
||
A 110M-param Llama-style decoder pre-trained from scratch on the BabyBabelLM
|
||
trilingual corpus (English, Dutch, Chinese), under the BabyLM 2026
|
||
MultiLingual track rules: **100M reference tokens, byte-premium adjusted**,
|
||
≤10 epochs.
|
||
|
||
This is the *baseline* zero-point of our ablation grid. Subsequent runs vary
|
||
the mixture allocation (loss-weighted, simultaneous-bilingual, typological-bridge
|
||
curriculum, register-controlled) on top of an identical scaffold. The matching
|
||
ablation paper is in preparation.
|
||
|
||
## Architecture
|
||
|
||
- Llama (HF `LlamaForCausalLM`) — RoPE, RMSNorm, SwiGLU, no biases, tied embeddings
|
||
- 12 layers · 768 hidden · 12 heads · 2048 FFN
|
||
- 1024 sequence length
|
||
- 110,119,680 parameters
|
||
|
||
## Tokenizer
|
||
|
||
Joint byte-level BPE, 32,768 vocab, trained on a balanced 50M-char sample from
|
||
each of EN/NL/ZH. The same tokenizer is shared across all three languages (see
|
||
the data card for why a joint tokenizer is required: ZH is 6.8% Latin script).
|
||
|
||
## Training
|
||
|
||
- **Data:** `BabyLM-community/babylm-eng` + `babylm-nld` + `babylm-zho`
|
||
(BabyBabelLM 2026 100M tier). Full corpora loaded in memory and shuffled
|
||
(the Hub layout is category-clustered; streaming with reasonable buffers
|
||
produces a biased sample).
|
||
- **Mixture:** byte-premium-uniform — equal share of *reference tokens* per
|
||
language (1/3 each), achieved by deficit-driven selection, not uniform doc
|
||
sampling (mean doc sizes differ across languages).
|
||
- **Optimizer:** AdamW (β₁=0.9, β₂=0.95, wd=0.1), lr 6e-4, cosine to 10%, 100-step warmup
|
||
- **Compute:** 4× NVIDIA A10G (23 GB), bf16, DDP, micro-batch 16 × grad-accum 2 (eff. batch 128 sequences = 131k tokens/step)
|
||
- **Tokens consumed at this checkpoint:** 100,000,000 byte-premium-adjusted reference tokens
|
||
- **Per-language epochs at this checkpoint:** ≈1.0 each (within the BabyLM ≤10-epoch cap)
|
||
|
||
## Revisions
|
||
|
||
The `chck_{N}M` revisions match the BabyLM eval pipeline's fast-eval naming:
|
||
|
||
```
|
||
chck_1M, chck_2M, ..., chck_9M, chck_10M, chck_20M, ..., chck_90M, chck_100M
|
||
```
|
||
|
||
Use `revision=chck_NM` to load any milestone. The default (`main`) is `chck_100M`.
|
||
|
||
## How to evaluate
|
||
|
||
```bash
|
||
git clone https://github.com/babylm-org/babylm-eval
|
||
cd babylm-eval/multilingual
|
||
bash scripts/zeroshot_model.sh --model_name Shamima/babylm-2026-multilingual-uniform-100M
|
||
bash scripts/zeroshot_model_fast_all.sh --model_name Shamima/babylm-2026-multilingual-uniform-100M
|
||
```
|
||
|
||
## Citation
|
||
|
||
```
|
||
@misc{babylm-2026-uniform,
|
||
title = {BabyLM 2026 MultiLingual baseline (byte-premium-uniform)},
|
||
author = {Hossain, Shamima},
|
||
year = {2026},
|
||
url = {https://huggingface.co/Shamima/babylm-2026-multilingual-uniform-100M}
|
||
}
|
||
```
|
||
|
||
Companion repo with audit, scaffold, and ablation configs:
|
||
https://github.com/silvererudite/bb-lm-challenge-sub
|