lm-distillation

Knowledge distillation

On real handwritten digits, a small student trained on a teacher's soft predictions over all data nearly matches the teacher and beats one trained on only a few hard labels.

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: Handwritten digits (UCI / scikit-learn)
• Type: real — public dataset
• Size / stats: 1,797 real 8×8 digit images (64-D), 10 classes; the plain student sees only 100 labels, the distilled one learns from the teacher's soft targets over all 1,257
• Split: 1,257 train / 540 test
• Source: scikit-learn load_digits (UCI Optical Recognition of Handwritten Digits)

Training config

Teacher: Adam (lr 2e-3), 800 steps. Student: Adam (lr 3e-3), 1500 steps; KL distillation at T=4 (loss ×16).

Evaluation results

Robustness (mean ± std over 5 seeds)

Single-run numbers above are one seed; this is the distribution over independent re-trains (honest variance — no cherry-picking).

Inference example

import torch, torch.nn as nn
teacher = nn.Sequential(nn.Linear(64,256), nn.ReLU(), nn.Linear(256,256), nn.ReLU(), nn.Linear(256,10))
teacher.load_state_dict(torch.load("teacher.pt", map_location="cpu")); teacher.eval()
# x: flattened 8x8 digit /16.0, shape (N,64)  ->  logits = teacher(x); pred = logits.argmax(-1)

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.

On 8×8 digits; the gain shrinks if the task is easy or the student is large; sensitive to temperature.

Failure cases

No benefit if the task is too easy or the student is already big enough; a wrong temperature washes out or over-sharpens the soft targets.

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_distillation.ipynb --output run.ipynb
# optional: override training length, e.g.  STEPS=2000  (or EPISODES=600)  before running

Files

• figure.png
• metrics.json
• seeds.png
• teacher.pt

License

Code & weights: MIT (this repository) — educational use encouraged.
Handwritten-digits data: UCI ML Repository via scikit-learn — CC BY 4.0.

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: Hinton, Vinyals & Dean, Distilling the Knowledge in a Neural Network, NeurIPS-W 2015.

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.