MonoWeb Models

Pretrained language models released alongside the paper:

The Role of Mixed-Language Documents for Multilingual Large Language Model Pretraining

Associated dataset: UCLNLP/monoweb-dataset

Model Details

All models are decoder-only transformers with 1.35B parameters, trained from scratch using the Llama-2 tokenizer (32K vocabulary). Architecture: 24 layers, hidden dimension 2048, 16 attention heads, context length 2048. Training was performed with Megatron-LM for ~143B tokens (34K steps).

Model Variants

Models are organized by language pair and training data configuration:

Folder	Language Pair	Training Data
ckpt_exp_en_de_baseline	English–German	FineWeb (full corpus, including bilingual docs)
ckpt_exp_en_de_monoweb	English–German	MonoWeb (bilingual docs removed)
ckpt_exp_en_de_onlyparallel	English–German	MonoWeb + parallel docs reintroduced
ckpt_exp_en_de_onlycodeswitch	English–German	MonoWeb + code-switching docs reintroduced
ckpt_exp_en_es_baseline	English–Spanish	FineWeb (full corpus, including bilingual docs)
ckpt_exp_en_es_monoweb	English–Spanish	MonoWeb (bilingual docs removed)
ckpt_exp_en_es_onlyparallel	English–Spanish	MonoWeb + parallel docs reintroduced
ckpt_exp_en_es_onlycodeswitch	English–Spanish	MonoWeb + code-switching docs reintroduced
ckpt_exp_en_fr_baseline	English–French	FineWeb (full corpus, including bilingual docs)
ckpt_exp_en_fr_monoweb	English–French	MonoWeb (bilingual docs removed)
ckpt_exp_en_fr_onlyparallel	English–French	MonoWeb + parallel docs reintroduced
ckpt_exp_en_fr_onlycodeswitch	English–French	MonoWeb + code-switching docs reintroduced

Each folder contains checkpoints saved every 2,000 steps from iter_2000 to iter_36000 (18 checkpoints per model).