nanoBeard-sloop-14M

nanoBeard ☠️

A tiny pirate-themed GPT trained from scratch on a piratized version of TinyStories, then SFT-tuned. Built as a learning project — closer to nanoGPT than to a production LM.

• Source code: https://github.com/younissk/pirate_llm
• Model on the Hub: https://huggingface.co/younissk/nanoBeard

Model details

Training

• Pretraining: TinyStories, piratized via a rule-based transform.
• SFT stage: stage=sft, iters=1400, val_loss=4.2816 (best 4.2485).
• Optimizer: AdamW, lr=2e-5, weight_decay=0.0, betas=(0.9, 0.95), grad_clip=1.0.
• LR schedule: linear warmup (50 steps) → cosine decay to min_lr=2e-6 over 1500 steps.
• Hardware/dtype: trained on cuda in bfloat16.

Files in the released repo

• model.safetensors — model weights.
• config.json — architecture config (load into training.config.Config).
• pirate_bpe.json — tokenizer (load with tokenizers.Tokenizer.from_file).
• training_metadata.json — full training config + metrics snapshot.
• banner.png — the banner above.

Usage

This model is not a transformers model — it uses the custom GPT class from this repo.

import json, torch
from huggingface_hub import hf_hub_download
from safetensors.torch import load_model
from tokenizers import Tokenizer

from training.config import Config
from training.model import GPT

repo = "younissk/nanoBeard"
cfg_path = hf_hub_download(repo, "config.json")
weights_path = hf_hub_download(repo, "model.safetensors")
tok_path = hf_hub_download(repo, "pirate_bpe.json")

cfg_dict = json.load(open(cfg_path))
cfg = Config(**{k: v for k, v in cfg_dict.items()
                if k in Config.__dataclass_fields__})
model = GPT(cfg).eval()
load_model(model, weights_path)

tok = Tokenizer.from_file(tok_path)
ids = torch.tensor([tok.encode("Once upon a time").ids])
with torch.no_grad():
    for _ in range(80):
        logits, _ = model(ids[:, -cfg.block_size:])
        next_id = torch.multinomial(torch.softmax(logits[:, -1] / 0.8, -1), 1)
        ids = torch.cat([ids, next_id], dim=1)
print(tok.decode(ids[0].tolist()))

Limitations

• Trained on a small synthetic corpus (TinyStories, piratized). Vocabulary, grammar, and world knowledge are extremely narrow.
• Short context window (256 tokens).
• No safety tuning. Outputs are pirate-flavored nonsense at best.
• Intended as an educational artifact, not a useful chat model.

Data

• Base corpus: TinyStories (CDLA-Sharing-1.0), transformed by the dataset/piratize.py script in this repo.