Processing Codeparrot Dataset Pipeline | Generated by AI

Home 2026.07

Done. Two new scripts committed:

Step Script Produces
1. Download scripts/download/download_codeparrot_clean.py .json.gz shards in datasets/codeparrot-clean/
2. Convert scripts/extract/convert_codeparrot_for_nanochat.py text-col parquet in datasets/codeparrot-clean-nanochat/
3. Tokenize scripts/extract/tokenize_github_code.py (existing) .npy token shards in datasets/codeparrot-clean-tok/

Smoke test (valid set, 1 shard, 61k Python files) — passed ✅

To run the full pipeline (all 53 train shards + 1 valid shard, ~12.7 GB compressed):

Step 1 — download:

python3.11 scripts/download/download_codeparrot_clean.py

Step 2 — convert to nanochat format:

python3.11 scripts/extract/convert_codeparrot_for_nanochat.py

Step 3 — tokenize:

python3.11 scripts/extract/tokenize_github_code.py \
  --input-dir /mnt/data/zz/datasets/codeparrot-clean-nanochat \
  --output-dir /mnt/data/zz/datasets/codeparrot-clean-tok

Timing estimate: step 1 ~15-20 min (14 MB/s observed), step 2 ~20-30 min, step 3 ~10-15 min. Total pipeline ~1 hr for 5.3M Python files.

The dataset is pure Python — deduplicated, filtered for quality (autogenerated flag, alpha fraction, line stats). Exactly what you need for a Python-focused nanoGPT or nanochat experiment.


Back Donate