Marco multilingual finetuning splits (c4)

Pre-built JSONL training splits for the Marco-MoE multilingual expansion experiments.

Contents

splits/ — document-budget tiers (~15 GB)

Fixed-prefix document counts per language (50k ⊂ 100k ⊂ 500k):

Language	Source corpus
dan_Latn	FineWeb2-HQ
fas_Arab	FineWeb2-HQ
swe_Latn	FineWeb2-HQ
tam_Taml	FineWeb-2
cat_Latn	FineWeb-2

Files: {lang}_{50k|100k|500k}.jsonl, {lang}_manifest.json, splits_summary.json

Generated by finetuning/data/prepare_data_splits.py.

token_budget_splits/ — token-budget tiers (~66 GB)

Fixed-seed shuffled splits at 1B / 1.5B / 2B raw tokenizer tokens:

Files: {lang}_{1Btok|1p5Btok|2Btok}.jsonl, {lang}_token_budget_manifest.json

Generated by finetuning/data/prepare_token_budget_splits.py (seed=42, Marco-Nano-Base tokenizer).

Usage

huggingface-cli download LiangYan3612/c4 \
  --repo-type dataset \
  --include "splits/**" \
  --local-dir finetuning/data

huggingface-cli download LiangYan3612/c4 \
  --repo-type dataset \
  --include "token_budget_splits/**" \
  --local-dir finetuning/data

Or use download_scripts/anvil/download_finetune_splits.sh from the multilingual_MoE repo.

Citation

Derived from FineWeb2-HQ and FineWeb-2. See project repo for experiment details.