MiniGPT-LATIN-21.15M

This is a custom 21.15M parameter autoregressive language model built from scratch in PyTorch and trained on a Latin text corpus.

Model Details

• Architecture: Custom MiniGPT (Pre-LN Transformer with RoPE)
• Parameters: 21.15M
• Tokenizer: 32K Latin BPE (SentencePiece)
• Context Length: 256 tokens
• Format: SafeTensors

Note: Because this is a custom nn.Module and not a standard transformers class, the Hugging Face Inference API widget will not work. You must download the weights and instantiate the custom class locally to run inference.