First 70B model released with all training epochs and data (huggingface.co)

Trillion Labs has released the first 70-billion parameter (70B) Korean-targeted large language model (LLM) with complete intermediate checkpoints covering all training epochs. This release, spanning models of 0.5B, 1.9B, 7B, and 70B parameters, offers researchers unprecedented access to progressive training snapshots captured at consistent token intervals—from roughly 20 billion tokens up to 160 billion tokens—enabling detailed analysis of training dynamics and model behavior at multiple scales. Notably, the smaller 0.5B and 1.9B checkpoints, initially used internally for system validation, are now publicly available as valuable references for studying early training phases in smaller LLMs. This transparency is significant for the AI/ML community as it provides open datasets to investigate how LLMs evolve over training, facilitating research on optimization, generalization, and scaling effects in Korean-language models. The checkpoints are accessible via Hugging Face, with easy integration through the Transformers library, allowing developers and researchers to load specific training stages using model revisions. By sharing loss curves and training configurations on their blog, Trillion Labs supports reproducibility and a deeper understanding of large-scale multilingual model training. This milestone sets a precedent for more granular, open releases of large models, promoting collaborative progress in LLM development and fine-tuning techniques.