tools/README.md

tools — HF dataset maintenance scripts

Bookkeeping copies of the scripts used to maintain the HuggingFace dataset ASKabalan/jax-fli-experiments (this repo is a local clone of it). These tools are not part of the dataset itself — they're versioned here for reference.

Run with uv so deps are available, e.g. uv run --with pyarrow --with huggingface_hub python3 tools/<script>.py.

reencode_parquet.py — probe / dry-run / re-encode parquet (dictionary + codec)

Losslessly re-encodes parquet to dictionary + a chosen codec to shrink it. Data is preserved bit-for-bit (verified per file: full-table equals + the huggingface schema metadata + the array column's Array2D extension metadata), so load_dataset features and config bundling are unaffected.

PATH may be a single .parquet file or a folder (recursed for **/*.parquet).

Three modes (positional):

mode	verifies	writes in place	what it's for
probe	no	no	just measure & print the size table
dryrun	yes	no	full safe pipeline, report would-change
reencode	yes	yes (if >2% win)	actually shrink files in place

reencode is atomic (temp + os.replace) and idempotent — files already dictionary-encoded in the target codec are skipped, so a run is resumable.

cd /home/wassim/Projects/NBody/jax-fli-experiments
uv run --with pyarrow python3 tools/reencode_parquet.py probe    06-cosmogrid-shells
uv run --with pyarrow python3 tools/reencode_parquet.py dryrun   06-cosmogrid-shells --codec ZSTD
uv run --with pyarrow python3 tools/reencode_parquet.py reencode 06-cosmogrid-shells

--codec defaults to SNAPPY (pyarrow's default; what the dataset already uses). Every mode ends with a per-file + per-folder + grand-total table of current / uncompressed / target / savings sizes.

After reencode, commit + push (one commit) from the clone:

git add -A && git commit -m "Re-encode parquet to dictionary+<codec> (lossless)"
git -c credential.helper='!f(){ echo username=ASKabalan; echo "password=$HF_TOKEN"; }; f' push origin main
git lfs prune   # reclaim local .git/lfs space from the old (now unreferenced) objects

squash_hf_history.py — reclaim HF storage after re-encoding

Re-encoding + push makes downloads smaller, but the old uncompressed LFS blobs stay referenced by the old commits, so HF keeps billing for them. This collapses the whole remote history into one commit so HF can GC the old blobs (storage drops to ~HEAD size).

Irreversible on the remote — destroys all commit history. The script always takes a local git bundle --all backup first (written next to the script), and only squashes with --yes.

# backup + dry-run
uv run --with huggingface_hub python3 tools/squash_hf_history.py REPO_ID LOCAL_CLONE
# actually squash
uv run --with huggingface_hub python3 tools/squash_hf_history.py REPO_ID LOCAL_CLONE --yes
# then re-align the local clone:
git -C LOCAL_CLONE fetch origin && git -C LOCAL_CLONE reset --hard origin/main

REPO_ID e.g. ASKabalan/jax-fli-experiments; LOCAL_CLONE is the local git clone path (used only for the safety bundle). Needs HF_TOKEN (write) in the env.