---
license: cc-by-4.0
task_categories:
  - text-generation
language:
  - en
tags:
  - space
  - space-situational-awareness
  - ontology
  - rdf
  - turtle
  - knowledge-graph
  - ontology-alignment
  - neurosymbolic
size_categories:
  - 1K<n<10K
---

# ssao-space-instruct

Instruction data teaching a language model to write **valid RDF Turtle in the
Space Situational Awareness Ontology** (SSAO) for real space objects, and to
judge proposed catalogue-to-ontology alignments using instance evidence.

1,245 examples: 1,072 train, 74 validation, 99 test. Built by
[Tesseract Academy](https://gov.tesseract.academy).

## The construction principle

**No example asserts anything a validator cannot check.** Every Turtle target is
*generated* from a real CelesTrak SATCAT record by applying a published, argued
SATCAT-to-SSAO alignment, so the ontology terms are correct by derivation rather
than by authorship. A **fail-closed vocabulary gate** refuses to emit the
dataset at all if any `ssao:` term in any target is missing from the ontology's
authoritative 554-term set.

That gate earned its place immediately: the first builder version invented
`ssao:hasOrbit` and `ssao:hasInclination`, which SSAO does not define (the real
names are `has_Orbit`, `has_Orbital_Inclination`). Shipping those would have
taught fluent hallucination in a domain whose ontology cannot reject it, which
is exactly the failure mode the accompanying study measures.

## Task families

| Family | n (approx) | What it teaches |
|---|---:|---|
| T1 record to Turtle | 900 | Catalogue fields to SSAO-typed Turtle with correct prefixes |
| T2 term lookup | 60 | Domain question to the correct SSAO class, with SSAO's definition |
| T3 alignment judgement | 55 | Accept or refuse a proposed correspondence, with argument and witness counts |
| T4 orbit regime | 200 | Orbital elements to regime class, with the deciding threshold cited |
| T5 refusal | 30 | Explicit refusal where the catalogue cannot determine the answer |

**T3 is the unusual one.** Positive rows come from a curated crosswalk; negative
rows come from its asserted non-mappings plus mappings that were empirically
refuted by counting counter-instances in the catalogue, with the witness counts
quoted as the justification. For example, refusing
`kg:RegimeGEORegion = ssao:Geostationary_Orbit` cites 1,007 of 1,734
geosynchronous-band residents whose inclination exceeds 5 degrees. We are not
aware of another public dataset covering alignment judgement with extensional
evidence.

## Sources and provenance

- **CelesTrak SATCAT**, snapshot 2026-07-29, pinned by sha256. 70,122 objects;
  T1 and T4 draw on a random sample. Data courtesy of T.S. Kelso.
- **Space Situational Awareness Ontology** by Robert J. Rovetto, in the copy
  vendored by the NASA `mission-viz` project (351 classes, 150 properties).
- **The alignment and the witness counts** come from
  [neurosymbolic-space-kg](https://github.com/fabio-rovai/neurosymbolic-space-kg),
  which contains the builder, the validator, the vocabulary gate and the
  measurements. Everything regenerates from source.

## Format

Chat-formatted JSONL, one object per line, with a fixed system prompt:

```json
{"messages": [
  {"role": "system", "content": "You are a space-domain knowledge engineer..."},
  {"role": "user", "content": "Express this catalogue record as SSAO Turtle...."},
  {"role": "assistant", "content": "@prefix ssao: <https://purl.org/space-ontology/> ...."}
]}
```

## Known limitations

- **Property coverage is deliberately narrow:** only the SSAO properties the
  catalogue can populate honestly (COSPAR number, country of origin, launch
  date, orbit, inclination, period, perigee, apogee). Most of SSAO's 150
  properties are untouched.
- **Status codes are operator-reported** and can be stale; T1 targets inherit
  that.
- **Derived orbit regimes rest on stated thresholds** (2000 km for LEO, a
  1400 to 1500 minute geosynchronous band, 5 degrees for geostationary
  candidacy). They are conventions, printed in the builder and adjustable; a
  sensitivity sweep of their influence is published in the repository.
- **The alignment judgements are a curated position**, reviewable and arguable,
  not ground truth.
- Roughly 72 percent of examples are one task family (T1), so a model trained
  on this alone will be strongest at record-to-Turtle.

## License

CC BY 4.0. SATCAT data courtesy of CelesTrak; SSAO by Robert J. Rovetto. Please
attribute both alongside this dataset.