Mangosteen

wannaphong 's Collections

updated 5 days ago

Mangosteen, a 47 billion-token Thai corpus built with a Thai-adapted pipeline, improves language model performance on Thai benchmarks.

Upvote

Mangosteen: An Open Thai Corpus for Language Model Pretraining

Paper • 2507.14664 • Published Jul 19 • 6
wannaphong/dolma_web1_raw

Viewer • Updated Jul 28 • 325M • 6

Note Raw data for Thai Dolma - Commoncrawl - Fineweb2
wannaphong/fineweb2_base

Viewer • Updated Jul 28 • 37.5M • 6

Note Fineweb2 LD
wannaphong/fineweb2_quality

Viewer • Updated Jul 28 • 19.3M • 6

Note + Quality filtering
wannaphong/fineweb2_quality_dedup

Viewer • Updated Jul 28 • 17.7M • 6

Note +Deduplication
wannaphong/fineweb2_dolma

Viewer • Updated Jun 8 • 16.9M

Note Fineweb2 by our pipeline (LD+ Quality filtering + Deduplication + Content filtering)
wannaphong/cc_base

Viewer • Updated Jul 28 • 165M • 6

Note Commoncrawl LD
wannaphong/cc_quality

Viewer • Updated Jul 28 • 34.5M • 6

Note + Quality filtering
wannaphong/cc_quality_dedup

Viewer • Updated Jul 28 • 25.4M • 6

Note +Deduplication
wannaphong/cc_dolma

Preview • Updated Jul 28 • 6

Note Commoncrawl by our pipeline (LD+ Quality filtering + Deduplication + Content filtering)
wannaphong/web-dataset

Viewer • Updated Apr 14 • 19.8M
wannaphong/thai-dolma-fasttext-model

Updated Jul 2

Note FastText Model
aisingapore/WangchanLION-Curated

Viewer • Updated Jul 22 • 402k • 246 • 3

Note Non-common crawl subset
aisingapore/WangchanLION-Web

Viewer • Updated Jul 22 • 19.8M • 87 • 3

Note common crawl subset
aisingapore/WangchanLION-v3

Text Generation • 8B • Updated Jul 22 • 93 • 3

Note CPT base model
aisingapore/WangchanLION-v3-IT

Text Generation • 8B • Updated Jul 22 • 59 • 3

Note SFT model

Upvote