Publicus/skillcenter-autoformal-span-cache

SkillCenter Intent source spans

This completed development subset contains 20,544 spans from 99 source documents, extracted from the first two corpus shards of Publicus/skillcenter-ir at revision 2cc11a73403d03c0679ffa909c893ef6a850048a. It is preparation data for IntentIR autoencoder experiments: source spans, context, provenance, and eligibility annotations. It contains no formal-logic targets or model weights, and it does not establish semantic correctness.

Coverage and selection

The upstream release contains 216,972 corpus rows in 53 shards. This extraction processed 8,192 rows in shards part-000000.parquet and part-000001.parquet. The source policy retained all 99 eligible documents: 8,056 rows required license review, and 37 additional rows failed the native source policy. Their contents are absent from this dataset. The remaining 51 shards, containing 208,780 rows, have not been processed by this span extraction run.

All 618,605 source characters (678,502 UTF-8 bytes) are represented exactly, including headings, whitespace, fenced code, and context-dependent material. The inventory has 1,485 candidate spans, 3,579 requiring context, 15,479 excluded from pairing, and one retained oversize span. Here, excluded means excluded from training-pair eligibility; these spans remain in the inventory to preserve the source document. Candidate eligibility is permission to attempt target extraction, not a reviewed training label.

Development split	Sources	Spans
train	70	14,991
validation	13	2,105
test	16	3,448

The splits preserve the local source-family assignments and duplicate fences. They originate from public upstream train shards and have already been used in development. They are not an independent or hidden benchmark test set. Repeated candidate text crossing partitions is retained but excluded from pairing.

Span extraction method

The shared sentence policy, uscode-whitespace-punctuation/v1, follows the whitespace normalization and punctuation-boundary approach used for the US Code autoformal span cache. The SkillCenter adapter, skillcenter-markdown-source-spans/v1, applies it inside Markdown blocks while preserving heading, list, preceding-scope, and code context. It retains complete oversize sentences rather than truncating their meaning. No LLM calls or source-content execution were used to extract these spans.

Loading

from datasets import load_dataset

spans = load_dataset("Publicus/skillcenter-autoformal-span-cache", "spans")
sources = load_dataset("Publicus/skillcenter-autoformal-span-cache", "sources")

For reproducibility, pass revision="<immutable Hub commit>" to both calls. Each Parquet file has at most 4,096 rows and uses Zstandard compression.

spans contains stable id, source_id, source_sha256, raw text, normalized_text, half-open start_char/end_char and UTF-8 start_byte/end_byte offsets, domain, split, status, reason, and eligible_for_pairing. Parse heading_context_json and block_context_json with json.loads to recover the original nested structures exactly.

sources contains each original instruction document, license expression, source URL, content digest, source identity and family JSON, and its preserved split. original_source_record_json retains the upstream metadata, title, bundle identity, source IDs, and text for attribution and replay; native_policy_json retains the recorded publication decision. source_shard, zero-based source_row_index, and document_index locate the source in the immutable upstream Parquet release.

The parent corpus derives from Tommysha/skillcenter-bundles revision f9dd4fec3c86d85ebf116c7408ac5ce602c418a1. manifest.json records source shard hashes, producer hashes, original export hashes, coverage, counts, and a complete release file allowlist with SHA-256 hashes. split-manifest.json preserves the original development partition evidence.

Verification and limits

After downloading a pinned snapshot, run:

python -m pip install pyarrow
python verify_release.py .

The verifier checks file hashes, row counts, source hashes, character/byte selectors, stable span IDs, exact document reconstruction, and split fences. The original native corpus loader was also replayed before packaging. This portable publication is a separate export format: its manifest does not replace the original local loader's absolute-path descriptor.

No proof admission, correct instruction-to-logic translation, generalization score, or supervisor benchmark improvement is implied. Extracted skill text remains untrusted data and is not authorization to execute embedded commands.

Attribution and licensing

Each document retains its source URL, original metadata, and declared license. The retained source declarations are MIT (85), Apache-2.0 (12), CC-BY-4.0 (1), and CC0-1.0 (1). The native source policy allowed training and publication for these records. These are inherited upstream declarations, not an independent audit of ownership. This packaging does not relicense the source material. See LICENSES.md and each source record for applicable terms.

Span extraction and packaging were performed with ipfs_datasets_py. Original source text and available notices are retained verbatim; normalized span text, segmentation, context annotations, and development partitions are derived additions.