metadata
datasets:
- infCapital/vnnews_corpus_100K
language:
- vi
Base Model: LLaMa2 7B Chat HF
- Extend vocab to 44,800 for better Vietnamese understanding
- Continual Pre-Train with >2B tokens Vietnamese
- Trainning profile: LoRa (rank=32, alpha=128, 16fp), 1 epoch, block size = 512. Takes 300GPU Hours x RXT4090 24GB
Can be better use for
- Futher training / Fine-tuning for Vietnamese tasks