75 lines
2.6 KiB
Markdown
75 lines
2.6 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 v2 (byte-premium-uniform, WSD)
|
|||
|
|
|
|||
|
|
Iteration v2 of the byte-premium-uniform trilingual baseline. Same
|
|||
|
|
architecture, same tokenizer, same mixture as v1
|
|||
|
|
(`Shamima/babylm-2026-multilingual-uniform-100M`), but trained with a
|
|||
|
|
**Warmup-Stable-Decay** schedule rather than cosine, on the same 100M
|
|||
|
|
ref-token budget at the same per-step compute. The motivation: v1's cosine
|
|||
|
|
schedule decayed to 10% of peak by ~90% through training while loss was
|
|||
|
|
still falling; WSD holds at peak and only decays in the last 25%, which lets
|
|||
|
|
more of the budget land at productive learning rates.
|
|||
|
|
|
|||
|
|
## 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 (same as v1; reused so the two are directly comparable)
|
|||
|
|
|
|||
|
|
## Training
|
|||
|
|
|
|||
|
|
- Data: BabyBabelLM 2026 100M tier (EN/NL/ZH); full corpora loaded in memory and shuffled
|
|||
|
|
- Mixture: byte-premium-uniform via deficit-driven selection (1/3 of *reference tokens* per language)
|
|||
|
|
- Optimiser: AdamW (β1=0.9, β2=0.95, wd=0.1)
|
|||
|
|
- LR: 6e-4 peak, **WSD schedule** (warmup 200 → constant peak → linear 25% decay tail to 6e-5)
|
|||
|
|
- 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: 100,016,896 byte-premium-adjusted reference tokens (= 1 epoch over the corpus)
|
|||
|
|
- Per-language epochs at this checkpoint: ~1.0 each (well within the BabyLM ≤10-epoch cap)
|
|||
|
|
|
|||
|
|
## Revisions
|
|||
|
|
|
|||
|
|
19 fast-eval branches: `chck_1M, chck_2M, …, chck_9M, chck_10M, chck_20M, …, chck_90M, chck_100M`.
|
|||
|
|
`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-v2
|
|||
|
|
bash scripts/zeroshot_model_fast_all.sh --model_name Shamima/babylm-2026-multilingual-uniform-100M-v2
|
|||
|
|
```
|
|||
|
|
|
|||
|
|
## Comparison vs v1
|
|||
|
|
|
|||
|
|
See https://github.com/silvererudite/bb-lm-challenge-sub for the iteration
|
|||
|
|
log, scaffold, and ablation configs.
|
|||
|
|
|
|||
|
|
## Citation
|
|||
|
|
|
|||
|
|
```
|
|||
|
|
@misc{babylm-2026-uniform-v2,
|
|||
|
|
title = {BabyLM 2026 MultiLingual baseline v2 (WSD schedule)},
|
|||
|
|
author = {Hossain, Shamima},
|
|||
|
|
year = {2026},
|
|||
|
|
url = {https://huggingface.co/Shamima/babylm-2026-multilingual-uniform-100M-v2}
|
|||
|
|
}
|
|||
|
|
```
|