Youniss

← All projects

A small pirate-themed GPT trained from scratch on a piratized TinyStories corpus and then SFT-tuned, with a versioned training codebase.

  • 2026
  • ai-ml
  • shipped

About nanoBeard

A from-scratch language model trained on a piratized version of TinyStories, then supervised fine-tuned, closer to nanoGPT than to a production LM, and honest about it. The first ship-class, Sloop, is published on Hugging Face; later classes (Brig, Frigate, Galleon) are configs in the same codebase.

A chat app where inference happens on the phone. Models ship as GGUF files pulled from Hugging Face on first use, and that download is the only network call the app makes, no accounts, no analytics, no server round-trip for a message.

Details

nanoBeard

Stack
  • Python
  • PyTorch
  • Hugging Face
  • Gradio
  • MkDocs
Tags
Links
Code
Training code99 starsiOS app3 stars
Models
sloop-14M3,407 downloadsgalleon-34Msloop-14M GGUF510 downloadsgalleon-34M GGUF574 downloadsfrigate-125M GGUF425 downloadsfrigate-360M GGUF1,295 downloads

More in ai-ml

See all
  • PaperNavigator
    • 2026
    • shipped

    Academic paper discovery tool that profiles a research question, snowballs through citations, filters with an LLM and generates a report.

  • DISCO
    • 2025
    • shipped

    CLIP-based classifier for detecting implicit suggestive imagery, trained on a hand-labelled fashion dataset and published on Hugging Face.

  • Embed2Image contrastive retrieval
    • 2025
    • experiment

    Audio-text retrieval experiments on Clotho comparing a PaSST+RoBERTa baseline against a pseudo-image ViT encoder head.

  • Falcon-Twig
    • 2025
    • shipped

    Fine-tuning run for Falcon H1 7B on tool calling that underperformed the base instruct model, written up in a post.