ML drills

Fill-in-the-blank coding drills for learning ML implementations.

setup

Put your Python implementations in drills/. Any top-level class or function becomes drillable.

transformer_drills/
  drills/
    attention.py
    transformer_block.py
    ...              ← add your own files here too
  drill.py
  drill_core.py
  eval_work.py

running

python drill.py

That's it. A browser window opens automatically at http://localhost:7234.

using the UI

Sidebar (left panel)

Control	What it does
Exercise	Pick a specific file + function/class, or leave on Random
Mask %	Percentage of code lines to blank out
Guidance hints	When on, blanks show # code here; when off, lines are completely empty
Generate	Creates a new exercise with the current settings

Editor (main panel)

• Blank lines have a blue left border and an inline input field
• Indentation is preserved — just type the code
• Tab / Shift+Tab moves between blanks
• Ctrl+Enter (or the Evaluate button) checks your answers

Results

• Correct blanks turn green
• Wrong blanks turn red with a strikethrough, and the correct answer appears on the line below — right in context
• Score shows in the sidebar

example for what's in drills/

File	Contents
attention.py	Masks, scaled dot-product attention, MHA, cross-attention
attention_variants.py	Multi-query attention, grouped-query attention, KV cache
positional.py	Sinusoidal encoding, learned embeddings, RoPE
normalization.py	LayerNorm, RMSNorm, BatchNorm
feedforward.py	FFN, GatedFFN (SwiGLU / GeGLU)
transformer_block.py	Pre-norm block, post-norm block, decoder block
architectures.py	TransformerEncoder, TransformerDecoder, EncoderDecoder, GPT, BERT
recurrent.py	RNNCell, LSTMCell, RNN, LSTM
losses_and_activations.py	Softmax, swish, GELU, SiLU, cross-entropy, focal loss
moe.py	Switch MoE (top-1), Soft MoE (top-k)
contrastive.py	InfoNCE, NT-Xent (SimCLR), CLIP loss, dual encoder
convolutions.py	Residual block, simple CNN (Conv2d), TCN (Conv1d + dilation)
classical_ml.py	Linear regression, logistic regression, precision/recall/F1, k-means
distributed.py	Data parallelism, model parallelism, pipeline parallelism
generative.py	Autoencoder, VAE, GAN, straight-through estimator, VQ-VAE

suggested workflow

1. Start with 5-10% mask and guidance on — learn the structure
2. Increase to 20–50% once the skeleton feels familiar
3. Turn guidance off — now you have to remember what's missing
4. Hit 100% mask on individual functions for blank-slate practice
5. Repeat the same function after 3 days, 1 week, 2 weeks