| # Banalata — banalata_2026-04-12_iter6900_val3.7835 | |
| Decoder-only transformer trained from scratch on public-domain Bengali | |
| literary text (bangla_sahitya dataset). | |
| ## Model details | |
| - Architecture : GPT (decoder-only), RoPE + RMSNorm + SwiGLU | |
| - Layers/Embd/Heads : 8 / 512 / 8 | |
| - Tokenizer : SentencePiece BPE, vocab=5000, trained on Bengali only | |
| - Training : iter=6900, best val loss=3.7835 | |
| ## Inference | |
| ```python | |
| python s05_generate.py --author "রবীন্দ্রনাথ ঠাকুর" | |
| python s05_generate.py --prompt "আকাশ ভরা সূর্য তারা" | |
| ``` | |
| ## License | |
| Trained on public-domain Bengali literature. | |
| Model weights: Apache 2.0. | |