NBR-1B: Brazilian Portuguese Language Model

NBR-1B is a 1.13 billion parameter language model trained from scratch for Brazilian Portuguese.

Model Details

Attribute Value
Parameters 1.13B
Architecture LLaMA-style (GQA, RMSNorm, SwiGLU, RoPE)
Hidden Size 2048
Layers 24
Attention Heads 16
KV Heads 4
Vocabulary 32,000 (BPE)
Context Length 2048
Training Tokens 3.12B
Final Loss ~2.8

Training Data

  • CulturaX PT (40%)
  • FineWeb-2 PT (52%)
  • mC4 PT (5%)
  • CC-News PT v2 (2%)
  • Books PT (1%)

Usage

This is a base model for text completion. Use with transformers library.

License

Apache 2.0

Downloads last month
23
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support

Datasets used to train elvisia/nbr-1b