Trillion Labs needed to build highly accurate Korean and multilingual LLMs, but the scarcity of high-quality Korean data and the CPU-heavy cost of language filtering, deduplication, and shuffling made large-scale data curation too slow for rapid experimentation.
Reported impact
GPU-accelerated curation cut processing time for 1.7 trillion tokens from 24 hours to 3.4 hours, improved Korean-language accuracy by 5% with 100B tokens of curated data, and delivered up to 10x savings in cost and energy compared with CPU-based processing.
Trillion Labs used NVIDIA NeMo Curator to build a GPU-accelerated curation pipeline for very large Korean-language training corpora. The workflow identifies Korean text, applies custom quality heuristics on top of NeMo Curator filters, and performs semantic and exact deduplication so only high-value content flows into model training. Running the pipeline on 8x H100 GPUs let Trillion Labs process 1.7 trillion tokens in hours instead of a full day on CPU infrastructure, which increased the number of training experiments the team could run and improved the accuracy of its Korean models. NeMo Curator's Dask-based distributed data engine also let Trillion Labs keep scaling curation and training without reworking the entire platform as model size and multilingual coverage expanded.
Raw Korean-language corpora are filtered and language-tagged in NeMo Curator, deduplicated and cleaned on H100 GPUs through Dask-based distributed jobs, then emitted as higher-quality datasets that Trillion Labs feeds into its multilingual LLM training runs.
5% accuracy improvement for the Korean language using 100B tokens of curated data.
Deduplication is one of the most time-consuming processes when handling a very large dataset.
Korean AI company Trillion Labs is advancing national AI development to overcome the challenge of building highly accurate LLMs for Korea.
Training time was reduced from 24 hours on CPU to 3.4 hours on 8x H100s for 1.7 trillion tokens of Korean language data.
Review how this reference was scored and trace each normalized claim back to its documented source.
Scoring methodology
Customer Reference Value measures usefulness as a customer reference. Evidence Confidence measures the quality and traceability of available evidence, not independent verification.
Traceability
Field coverage shows which facts each source supports. Dates are shown only when documented by the source.
The public company profile reports '임직원 수(2026-08) 24명' for the South Korea-based company.
Related
NVIDIA Sample Sep 22, 2026 Sluicebox Sluicebox needed to turn fragmented bills of materials, supplier records, PDFs, CAD files, and emissions inventories into timely product carbon assessments, but manual life-cycle-analysis work was slow, expensive, and… 51-80% lower inference costs Lucy · NVIDIA Nemotron 3 Ultra · carbon graph NVIDIA Sample Sep 15, 2026 Lambda Lambda needed to raise token throughput and improve AI-cloud economics, but fixed rack and facility power budgets were leaving GPU capacity stranded and limiting how many nodes could run safely at once. +17% Hybrid Training Throughput Dynamic Power Software · NVIDIA DSX AI Factory Platform · NVIDIA DSX MaxLPS NVIDIA Sample Sep 1, 2026 Onetera Onetera needed to help more than 90 municipalities turn fragmented regulations, zoning maps, permit rules, and legacy documents into fast resident-facing services without imposing heavy IT overhead or expensive manual… 1.7x increase in data throughput GPT OSS 20B · Government Intelligence Agent · LangGraph 