Datasetpublic

atenareply/asterion-cpt-corpus

Asterion CPT Source Corpus What's inside — 297,185 synthetic technical documents about the fictional Asterion Space Operations fleet (24 satellites: EO/COMM/SCI/TD): 7 doc_types × 12 topics, ~2.5 GB of text, ~1.62B Gemma-4 tokens (measured mean 5,413 tokens/doc on a stratified 2k sample, 2026-07-03). Where it comes from — Fully synthetic — generated incrementally in 1,000-doc shards by the Asterion corpus generation pipeline (see noval-corp/docs/asterion-corpus-plan.md) from a… See the full description on the dataset page: https://huggingface.co/datasets/atenareply/asterion-cpt-corpus.

sourceHugging Faceapache-2.0updated 3mo agoView on Hugging Face

0likes347downloads

Dataset Card

Asterion CPT Source Corpus

What's inside — 297,185 synthetic technical documents about the fictional Asterion Space Operations fleet (24 satellites: EO/COMM/SCI/TD): 7 doc_types × 12 topics, ~2.5 GB of text, ~1.62B Gemma-4 tokens (measured mean 5,413 tokens/doc on a stratified 2k sample, 2026-07-03).

Where it comes from — Fully synthetic — generated incrementally in 1,000-doc shards by the Asterion corpus generation pipeline (see noval-corp/docs/asterion-corpus-plan.md) from a shared fleet/topic specification. No real-world sources; the domain is invented.

How it was built / modified — Raw source resource: no dedup, no held-out, train split only. Downstream CPT builders carve a stratified validation split and run exact + MinHash/LSH near-dedup. Generation COMPLETE 2026-07-03 (final partial shard 297); all builds pin revision 7f0c3236.

Overview

—Role: SOURCE corpus — Asterion round family (Gemma-4-12B full, LFM2.5 sampled)
—Lineage: fleet/topic spec → generation pipeline → this → buildasterioncorpus.py (Gemma round, full) / buildasterioncorpus_lfm2.py (LFM2.5 round, 55M-token stratified sample)

Schema

Column	Meaning
`text`	one document per row
`doc_type`	one of 7 (anomalyincidentreport, designreviewmemo, engineeringnote, missionopslog, subsystemstatusreport, technicalprocedure, telemetryanalysisreport)
`topic`	one of 12 fleet-operations topics
`word_count`	document length in words
`idx`	generation index

Splits

Split	Rows	Note
train	297,185	raw generation output (no held-out — carved downstream)

Used by

—buildasterioncorpus.py → atenareply/asterion-training-corpus-gemma4
—buildasterioncorpus_lfm2.py → atenareply/asterion-training-corpus-lfm2
—evalheadroomasterion.py / evalheadroomlfm2_asterion.py (pre/post-CPT PPL)

Limitations

—Formulaic templated style: perplexity is low even for models never trained on it — headroom checks matter more than raw loss.
—Shards were committed incrementally: ALWAYS pin a revision (builds on floating main are not reproducible).
—Fictional domain — no real spacecraft data anywhere in the corpus.

Part of the **noval-corp** fictional-domain training family (Orbital Mining Corporation / Mars Express; Asterion round for Gemma-4). Card generated by `noval-corp/scripts/gendatasetcards.py` — see `noval-corp/NOTE.md` for the full project log.