pharo-llm/pharo-reranker-dataset

Pharo reranker dataset: frozen corpus and runtime snapshot

This snapshot preserves 3,441,607 ranking examples from 732 Pharo packages, the original mining image and VM, frozen benchmark source, the pipeline source, and the existing experiment selection. It was prepared on 2026-10-08 from heuristics-vs-llm commit 813d45fd951bc861e905ee8122e7a7172efa12be. No pipeline code was changed to create this snapshot.

Contents

Path	Contents
mining/corpus/all.jsonl.gz	Losslessly compressed original JSONL, in its original row order
mining/corpus/manifest.json	Original row counts, package counts and source/image/corpus SHA-256 values
mining/corpus/packages.json	Original eligible package list
artifacts/pharo-image-macos-arm64.zip	mining/image/: original Pharo image, changes, sources, VM and launchers
artifacts/workspace.zip	All Git-tracked pipeline files and tests; mining/repository/ and image-ready marker; current experiment/ image, frozen source, split, model names and incomplete run metadata
snapshot.json	Source revision, environment observations, per-member hashes, file modes, exclusions and limitations
SHA256SUMS	Checksums of every published payload and metadata file except this checksum list itself

The corpus dataset split is the entire mined pool, not a training partition. Preserve the package holdouts below when training and evaluating.

Saved selection

Seed: 42. After reserving the existing 5 benchmark packages, the remaining packages use an 80/10/10 train/validation/test split.

Partition	Packages	Examples
Training	581	2,685,649
Validation	73	433,973
Test	73	307,853
Benchmark	5	14,132

experiment/split.json is the authoritative four-way split. experiment/benchmark-selection.json preserves the Pharo-side selection. All four sets of package names are disjoint. Benchmark rows must be excluded from model development. The existing incomplete reranker-run.* is included for reference; start a fresh run rather than trying to resume it.

Restore the same inputs

Use a new empty directory, with at least 32 GB of free disk for extraction and training partitions, plus additional space for models and package workers. Each package worker needs its own writable image. The supplied VM runs on macOS ARM64; Linux/OAR needs a compatible Linux VM and launcher before execution. Do not run the macOS VM on Linux or replace the original image by a fresh download.

Install the Hugging Face CLI separately if needed. Copy this snapshot's full commit SHA from the repository's commit history into REVISION below. Always use that fixed SHA for later downloads instead of main.

REVISION=REPLACE_WITH_FULL_DATASET_COMMIT_SHA
hf download pharo-llm/pharo-reranker-dataset \
  --repo-type dataset --revision "$REVISION" --local-dir pharo-snapshot
cd pharo-snapshot

# macOS (use sha256sum -c SHA256SUMS on Linux):
shasum -a 256 -c SHA256SUMS

unzip artifacts/workspace.zip
unzip artifacts/pharo-image-macos-arm64.zip
gzip -dk mining/corpus/all.jsonl.gz
chmod +x pipeline-benchmarks.sh pipeline-mine-training-data.sh \
  mining/image/pharo mining/image/pharo-ui \
  mining/image/pharo-vm/Pharo.app/Contents/MacOS/Pharo \
  experiment/image/pharo experiment/image/pharo-ui \
  experiment/image/pharo-vm/Pharo.app/Contents/MacOS/Pharo

python3 - <<'PY'
from pathlib import Path
import sys
sys.path.insert(0, 'scripts')
from ranking_corpus import verify, load_manifest
verify(Path('mining/corpus'), image=Path('mining/image/Pharo.image'),
       repository=Path('mining/repository'), split_path=Path('experiment/split.json'))
load_manifest(Path('mining/corpus'), image=Path('experiment/image/Pharo.image'),
              repository=Path('experiment/repository'))
print('Corpus, original images, source and saved split verified')
PY

# Choose concurrency for the available RAM/CPU/GPU; 8 is an example.
# This uses the restored selection of 5 benchmark packages.
BENCHMARK_JOBS=8 ./pipeline-benchmarks.sh

Do not set a different BENCHMARK_PACKAGE_COUNT, seed or split ratio when reusing this experiment. The pipeline will reject settings that conflict with the saved split. The archived root README.md documents the existing pipeline in detail. Extracting workspace.zip replaces this dataset card with that original README; the dataset card remains available at the pinned Hugging Face revision.

Reproducibility boundary

The snapshot preserves exact input bytes, package selections and source code. The original corpus SHA-256 is 7d93f4b3b38bf9fc77c13992066f807171aceaeeee19e8f71ba8e0e0a49b91bd. The original Pharo image SHA-256 is 4df757c743e18c3dcc191a5fd179be8b7cbd60790501360e7817352deb1d7a96.

It does not establish bit-identical model training, live LLM responses or benchmark timings. The workspace has no completed trained model or benchmark results, and does not contain Ollama weights or immutable Ollama model digests. Python dependencies use version ranges; observed local versions in snapshot.json are provenance only, not a tested dependency lock. Hardware, runtime versions, randomness and concurrency can affect results. Automatic pinned downloading, Linux runtime support and further determinism/parallelism changes are separate pipeline work.

Local credentials and all pharo-local/ state, Python caches, .DS_Store, the Git database and the obsolete previous-downloaded-image.* are excluded. The original image itself is unchanged. Third-party runtime and source components retain their upstream licenses; the repository's MIT metadata does not replace those licenses.