NumFaLM 3B is a bilingual language model trained in Thai and English. The architecture model is Llama model that pretraining from scratch. It was built to open source AI and research for bilingual language models and improve small language models. We released the training script and train datasets so you can research the training and datasets.

We fork EasyLM and added training by HuggingFace datasets, but HuggingFace was down many times during the time we trained the model, so we can train just one epoch. The model trained one epoch.

Research supported with Cloud TPUs from Google's TPU Research Cloud (TRC). We use TPU4-64 for training model about 4 days / 1 epoch.

Thank you TPU Research Cloud and EasyLM project! We use EasyLM for pretraining model.