A MoonBit dataset grouping, splitting, and leakage audit library for reproducible machine learning workflows.

dataset

split

leakage

research

Download zip

Author

AlexenderSokolov

Version

0.3.0

License

Apache-2.0

Repository

github.com/AlexenderSokolov/moonsplit

Last updated

7 days ago

Downloads

Dependencies

moonbitlang/x0.4.46

#MoonSplit

MoonSplit is a MoonBit CLI and library for leakage-aware dataset splitting and independent audit. Schema v2 combines declared exact keys, Unicode text near-duplicates, and source-scoped interval overlap into indivisible leakage components; it also provides explicit time-forward validation with per-record exclusion reasons.

The release line is 0.3.0. docs/acceptance.md separates locally verified evidence from the tag, CI, registry, and clean-install evidence recorded when the release is published. Official hackathon acceptance remains a decision of the organizers.

#Features

Schema v1-compatible exact-key grouping with transitive closure.
Schema v2 leakage rules: exact_key, Unicode character q-gram text_jaccard, and same-source half-open interval_overlap.
Deterministic grouped holdout/K-fold allocation with a bounded local component-move optimization that never worsens greedy initialization.
Time-forward validation windows with train, validation, and explicit excluded reasons.
Independent audit rebuilt from raw records, spec, and assignment/role manifests.
Canonical spec normalization and no-write validation preflight.
Stable JSON, JSONL, and Markdown exports, including label-aware summaries and K-fold membership manifests.
wasm-gc and JS targets.

#Installation

Run the checked-out source to reproduce this version:

git clone https://github.com/AlexenderSokolov/moonsplit.git
cd moonsplit

After Mooncakes publication, the exact versioned installation command is recorded only after a clean-directory installation test and is linked from docs/acceptance.md. Do not infer package availability from the source version alone.

The examples/ directory contains ready-made JSON, JSONL, and split specs.

#Quick Start

./run_check.ps1
./run_acceptance.ps1
./run_demo.ps1

./run_check.sh
./run_acceptance.sh
./run_demo.sh
./run_bench.sh

#Commands

moon run cmd/moonsplit -- plan --data examples/voice_records.jsonl --spec examples/holdout_spec.json --out examples/demo_run
moon run cmd/moonsplit -- audit --data examples/voice_records.jsonl --spec examples/holdout_spec.json --assignments examples/demo_run/assignments.jsonl
moon run cmd/moonsplit -- audit --data examples/voice_records.jsonl --spec examples/holdout_spec.json --assignments examples/demo_run/assignments.jsonl --report json
moon run cmd/moonsplit -- normalize-spec --spec examples/holdout_spec.json
moon run cmd/moonsplit -- validate --data examples/voice_records.jsonl --spec examples/holdout_spec.json --format jsonl
moon run cmd/moonsplit -- plan --data examples/v2_rule_records.jsonl --spec examples/v2_rule_spec.json --out artifacts/v2_plan --format jsonl
moon run cmd/moonsplit -- audit --data examples/v2_rule_records.jsonl --spec examples/v2_rule_spec.json --assignments artifacts/v2_plan/assignments.jsonl --format jsonl --report json
moon run cmd/moonsplit -- plan --data examples/time_records.jsonl --spec examples/time_forward_spec.json --out artifacts/time_plan --format jsonl
moon run cmd/moonsplit -- audit --data examples/time_records.jsonl --spec examples/time_forward_spec.json --roles artifacts/time_plan/roles.jsonl --format jsonl --report json
moon run cmd/moonsplit -- bench --mode interval --records 100000
moon run cmd/moonsplit -- bench --mode text --records 10000
moon run cmd/moonsplit -- demo --out examples/demo_run
moon run cmd/moonsplit -- --version
moon run cmd/moonsplit -- smoke --data examples/voice_records.jsonl

The output directory must not already exist. The parent directory must exist. Exit codes are:

0: success
2: usage, input, or read error
3: audit failure
4: output write error

smoke reads one input file and prints its character count for quick file-system checks.

Schema v1 holdout plans write five files and v1 K-fold plans write six. Schema v2 grouped holdout writes seven (evidence.jsonl and audit.json are added), grouped K-fold writes eight (plus folds.jsonl), and time_forward writes eight role/audit files. Existing v1 APIs and input schemas remain available. See docs/output.md and docs/api.md.

The output files are documented in docs/output.md.

#Scope

Text Jaccard detects character-overlap near duplicates; it does not claim semantic paraphrase, audio, image, or video similarity. Interval overlap only compares records sharing the declared source key and treats intervals as [start, end), so touching endpoints do not match. Time coordinates must use the explicitly declared interval, unit, and epoch; source-local character offsets are not global time. MoonSplit does not train models, schedule jobs, manage databases, or claim that a dataset is completely leak-free.

#v2 record and specification shape

V2 records retain id, optional label, and keys; they add optional text and named intervals:

{"id":"r1","keys":{"source":"doc-1"},"text":"example text","intervals":{"segment":{"start":0,"end":20}}}

The default text rule uses Unicode 3-gram sets and threshold_bp: 8000. It folds ASCII case and repeated whitespace, preserves Chinese characters and punctuation, and treats equality at the threshold as a match. Missing or empty normalized text is an input error when that rule is enabled. A dense corpus may require quadratic candidate comparisons; MoonSplit does not silently cap them, and rejects text_count_overflow before its portable Int evidence counters could wrap. Memory capacity remains a practical limit for unusually dense corpora.

time_forward assigns each record in every validation window in this order: complete-window validation; other members of a validation component as excluded/component_conflict; historical training ending no later than validation_start - gap; then gap, future, or boundary_overlap exclusions. Active train and validation sets must both be non-empty.

Site source Report issue Packages Build queue Skills Statistics