Download fetch/fetch_wikitext2.sh from TitanMamo10/k2-quant-eval: direct link, hf CLI and curl.
- Browser
- Download file 2.3 kB
-
https://huggingface.co/datasets/TitanMamo10/k2-quant-eval/resolve/main/fetch/fetch_wikitext2.sh
- Command line
-
hf download hf://datasets/TitanMamo10/k2-quant-eval/fetch/fetch_wikitext2.sh
-
curl -L -o fetch_wikitext2.sh https://huggingface.co/datasets/TitanMamo10/k2-quant-eval/resolve/main/fetch/fetch_wikitext2.sh
2.3 kB
| # fetch_wikitext2.sh [DEST] -- get WikiText-2 raw (CC BY-SA 4.0) for calibration + PPL. | |
| # | |
| # DEST/wikitext-2-raw/wiki.train.raw calibration + SC + trace (the pipeline's WIKI) | |
| # DEST/wikitext-2-raw/wiki.test.raw PPL eval (the pipeline's F) | |
| # | |
| # We do NOT redistribute this text (see ../LICENSE-DATA.md). This script does, from the | |
| # canonical Hub dataset repo Salesforce/wikitext, split wikitext-2-raw-v1. | |
| # | |
| # Requires: python3 with huggingface_hub + pyarrow (both in `pip install datasets`). | |
| set -eu | |
| DEST=${1:-.} | |
| DEST=$(mkdir -p "$DEST" && cd "$DEST" && pwd) | |
| python3 - "$DEST" <<'PY' | |
| import os, sys | |
| from huggingface_hub import hf_hub_download | |
| import pyarrow.parquet as pq | |
| dest = os.path.join(sys.argv[1], "wikitext-2-raw") | |
| os.makedirs(dest, exist_ok=True) | |
| for split, fn in (("train", "wiki.train.raw"), ("test", "wiki.test.raw")): | |
| p = hf_hub_download("Salesforce/wikitext", f"wikitext-2-raw-v1/{split}-00000-of-00001.parquet", | |
| repo_type="dataset") | |
| rows = pq.read_table(p).column("text").to_pylist() | |
| # NOTE: the Hub parquet is a whitespace-normalised conversion of the original release | |
| # tarball: per-line leading/trailing spaces and runs of blank lines are collapsed, so this | |
| # file is ~0.10% longer than our .raw copy (blank lines are preserved as separate rows). Same corpus, same content, slightly | |
| # different byte layout -> chunk boundaries (and therefore the last decimals of PPL) differ. | |
| out = os.path.join(dest, fn) | |
| with open(out, "w", encoding="utf-8") as f: | |
| f.write("\n".join(rows)) | |
| print(f"{out} {os.path.getsize(out)} bytes {len(rows)} rows") | |
| PY | |
| cat <<'EOF' | |
| Reference: this is the corpus we used, but our exact files are | |
| wiki.train.raw 10940747 bytes sha256 6707892fa3788b5ab9ed78ab5ff37d9fe825f6011a2ad4fcd6a6d467f0e7da57 | |
| wiki.test.raw 1290590 bytes sha256 173c87a53759e0201f33e0ccf978e510c2042d7f2cb78229d9a50d79b9e7dd08 | |
| (the hashes pin the upstream release layout; the parquet above will NOT match them). | |
| For a bit-identical PPL reproduction, ask us for those two files (CC BY-SA 4.0) or fetch the | |
| original release zip, which is currently unreachable from this endpoint (the Salesforce S3 | |
| bucket redirects to a host with a broken certificate, 2026-10-27). | |
| EOF | |