nanotransformer 0.1.0

Tiny causal language model (context_len=3) with training and evaluation tools.

NanoTransformer

Stage-wise implementation of a tiny causal LM (context_len=3, ~2.1M params).

PyPI install

pip install nanotransformer

CLI entry points (you supply a config JSON):

nanotransformer-train --config /path/to/config.json
nanotransformer-eval --config /path/to/config.json --checkpoint /path/to/checkpoint.pt
nanotransformer-baseline --config /path/to/config.json --k 0.1
nanotransformer-latency --config /path/to/config.json --checkpoint /path/to/checkpoint.pt

Stage 1: Environment

python -m venv venv
source venv/bin/activate
python scripts/install_deps.py

Stage 2: Tokenizer and data

1. Download a 50MB English corpus (optional helper):

python scripts/download_corpus.py --out data/corpus.txt --target-mb 50

2. Or place your UTF-8 corpus at data/corpus.txt
3. Train BPE tokenizer:

python scripts/train_tokenizer.py --corpus data/corpus.txt --vocab-size 10000 --out artifacts/tokenizer.json

4. Build token ids and splits:

python scripts/prepare_data.py --corpus data/corpus.txt --tokenizer artifacts/tokenizer.json --out artifacts/data_ids.pt --splits-out artifacts/splits.pt

Stage 3: Train

python train.py --config configs/base.json

Stage 4: Evaluate

python eval.py --config configs/base.json --checkpoint checkpoints/best.pt

Stage 5: Trigram baseline (add-k smoothing)

python baseline_trigram.py --config configs/base.json --k 0.1

Stage 6: Latency

python measure_latency.py --config configs/base.json --checkpoint checkpoints/best.pt --iters 200 --warmup 10

Notes

• Adjust hyperparameters in configs/base.json
• Outputs: artifacts/ (tokenizer, data, metrics) and checkpoints/ (model)