wannaphong / NumFaLM

NumFaLM (น้ำฟ้า) is a bilingual language model trained on Thai and English.

Geek Repo:Geek Repo

Github PK Tool:Github PK Tool

NumFaLM

NumFaLM (น้ำฟ้า) is a bilingual language model trained in Thai and English.

NumFaLM 3B

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 (april 2024), so we can train just one epoch. The model trained one epoch.

Steps

Datasets

For English, English is bigger than all Thai datasets, so we select the high quality dataset only. We choose cosmopedia dataset that wants to replicate Phi-1.5, minipile is a 6GB subset of the deduplicated The Pile corpus, and goodwiki is a 179 million token dataset of English Wikipedia articles collected on September 4, 2023.

For Thai, we create many datasets that are part of the Thai corpus. To build a bilingual language model, we add a Thai-English parallel corpus.

We added a Python coding corpus to make the language model have logic coding.

You can see the list dataset at datasets/.

Acknowledgements

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 the pretraining model.

About

NumFaLM (น้ำฟ้า) is a bilingual language model trained on Thai and English.

License:Apache License 2.0


Languages

Language:Jupyter Notebook 60.4%Language:Python 39.6%