nanoGPT — Tiny Shakespeare

A character-level GPT (decoder-only transformer) trained from scratch; best-checkpoint by validation loss.

Trained from scratch in Ropedia Academy — an interactive, bilingual course on embodied & spatial AI. Educational model: small and quick to train; the value is the method and a reproducible pipeline, not a leaderboard score. Try it live in the Ropedia demos Space.

At a glance

Dataset

• Name: Tiny Shakespeare
• Type: real (public-domain text)
• Size / stats: 1,115,394 characters (~1.1 MB); 65-character vocabulary
• Split: 90% train / 10% val
• Source: https://github.com/karpathy/char-rnn (data/tinyshakespeare)

Training config

AdamW (lr 3e-4, weight-decay 0.1), 3000 steps; char-level decoder-only transformer; best by val loss.

Evaluation results

Sample output

Ciegvate tumpot of Bad'ers
We narvervy sures toak hasing more,
This hous mad the dide to to the for for hard with to
IsSeet if love true;
Mught and how fath quear uppose? City hat.
My and main thou but staltany; him comblead.

LEUMIEN:
Charth eyet not, bath brans yoer
Where shat? I'll at har comen mort, thou gene.

FROMOK:
I shall even Romen of joysed
You kind indswaul'd with thou bakeng, with mell.

DUCIO:
My shall stalk you fall hear:
Mant shall In in brothere'y! prancer, best worde houm's afd it.

UCIIO!

GLINIUS:
Rell w, it?

Go ELIA:
Where.

HED VOLLOONT:
More, now he now of nather'd nev

Inference example

import torch
sd = torch.load("gpt.pt", map_location="cpu")   # decoder-only transformer weights
# Rebuild the GPT class + char vocab from the notebook (see "Reproduce"), then:
# model.load_state_dict(sd); model.eval()
# print(decode(model.generate(torch.zeros((1,1), dtype=torch.long), 500)[0].tolist()))
# (A pre-generated sample is included as sample.txt.)

Limitations

Educational scale. Trained quickly on CPU on small or synthetic data, so absolute numbers are not competitive with production systems — the value is the method and a reproducible pipeline. No large-scale data, no hyperparameter sweep, and no multi-seed variance is reported. Not for production use.

Character-level and tiny → no long-range coherence; it imitates Shakespearean style, not meaning.

Failure cases

Repetition, invented words, and no factual grounding — a small char-LM memorizes style, not content.

Reproduce / train your own

One click: open the notebook in Colab → Runtime → GPU → Run all, then run its Publish to the Hugging Face Hub cell.

From a shell:

git clone https://github.com/ChaoYue0307/ropedia-academy.git && cd ropedia-academy
pip install torch numpy matplotlib scikit-learn scikit-image gymnasium
jupyter nbconvert --to notebook --execute notebooks/training/LM_nanogpt_pretrain.ipynb --output run.ipynb
# optional: override training length, e.g.  STEPS=2000  (or EPISODES=600)  before running

Files

• config.json
• figure.png
• metrics.json
• model.pt
• sample.txt

License

Code & weights: MIT (this repository) — educational use encouraged.
Text: Tiny Shakespeare — public domain.

Citation

If you use this model or the course materials, please cite:

@misc{ropedia_academy,
  title  = {Ropedia Academy: an interactive course on embodied & spatial AI},
  author = {Ropedia Academy},
  year   = {2026},
  howpublished = {\url{https://chaoyue0307.github.io/ropedia-academy/}}
}

Method / original work: Vaswani et al., Attention Is All You Need, NeurIPS 2017; Karpathy, nanoGPT.

Related assets

• 🚀 Live demos: https://huggingface.co/spaces/cy0307/ropedia-demos
• 🤗 All trained models + collection: https://huggingface.co/cy0307
• 📚 Course & all labs: https://chaoyue0307.github.io/ropedia-academy/ · Labs tab
• 💻 Source / notebooks: github.com/ChaoYue0307/ropedia-academy

---

Part of the Ropedia Academy trained-model collection. Contributions & issues welcome on GitHub.