amDANIEL2024/gptmdogo-bible

Char-level GPT (Bible text)

Custom-architecture GPT trained from scratch on a character-level tokenizer.

• Training step: 4000
• Best val loss: 1.1844894886016846
• Block size at save: 256

This is not a transformers-compatible model. To load it, download gpt_best.pt and use the same GPT class definition from the training notebook:

ckpt = torch.load("gpt_best.pt", map_location="cpu", weights_only=False)
model = GPT(ckpt["config"])
model.load_state_dict(ckpt["model_state_dict"])
stoi, itos = ckpt["stoi"], ckpt["itos"]