Starhorse Access

StarHorse Access — SHboost-2024 & SH21 EDR3

When to Use

Use this skill for ANY StarHorse-related data work. Two products:

1. SHboost-2024 — XGBoost-approximated posteriors, public Parquet on S3
2. SH21 EDR3 — Full Bayesian StarHorse posteriors for Gaia EDR3, via gaia.aip.de TAP

Data Papers & Acknowledgment

Cite the source papers in any work using these products:

• SH21 EDR3 — Anders, Khalatyan, Queiroz et al. 2022, A&A 658, A91.
• SHboost-2024 — Khalatyan, Anders, Chiappini et al. 2024, A&A 691, A98.

These are access instructions only; the catalogs remain the authors' work under their own terms. Follow the gaia.aip.de data-acknowledgment policy when publishing.

Quick Decision

• **Nearby stars (= 0.95` for flag-cleaned sample

Plotting: CMD Hexbin (reproduces paper Figure 5)

fig, ax = plt.subplots(figsize=(7, 7))
hb = ax.hexbin(df['bprp0'], df['mg0'], gridsize=80, cmap='jet',
               norm=mpl.colors.LogNorm(), mincnt=1, edgecolors='none')
ax.invert_yaxis()
ax.set_xlabel(r'$G_{BP} - G_{RP}$')
ax.set_ylabel(r'$M_G$')

Pitfalls

• Pandas float32 boolean chaining: (df['col'] > 0) & (df['col2'] < 1) — wrap EVERY comparison in parentheses for float32 columns.
• ADQL vs PostgreSQL: get_one_query() uses language='postgresql'. ADQL via run_sync() is simpler for routine queries.
• run_sync() returns pyvo TAPResults, NOT DataFrame. Use .to_qtable().to_pandas() or pd.DataFrame({c: data[c] for c in data.colnames}). Column names are in .colnames, NOT .fieldnames.
• Run in foreground. PostgreSQL TAP jobs on gaia.aip.de typically complete in ~17 s for 200 k rows (verified 2026-06-30), but use timeout=300 to be safe. If you launch via terminal(background=true), the job keeps running after the session ends and creates stray completion nudges. Run the query script in foreground.
• Data quality: ~98.6 % valid rows. A LIMIT 200000 query returns ~197 k rows with non-null bprp0/mg0. The ~1.4 % nulls are stars lacking BPRP or MG photometry. Always check .notna() before plotting.
• submit_job returns immediately; job.run() starts execution. You must call both.
• Timeout on job.wait() is per-wait. If the first 10 s wait times out (job still processing), call job.wait() again in a loop or set a longer timeout.
• No LIMIT syntax issue. PostgreSQL dialect supports LIMIT natively. ADQL requires TOP N or WHERE random_index < N.
• ADQL async endpoint may return 404. If you see /async/phase 404 errors, switch to ADQL via tap_service.search() instead.
• shphotoflag and shoutflag are VARCHAR. Always compare with string literals.
• All float columns are float32. Convert to float64 before chained boolean masks.
• fidelity is 0–1 continuous. Use for quality cuts.

Verification

• Data source and access path explicitly recorded.
• Key columns and caveats documented.
• SH21 schema verified 2026-06-30 against live gaia.aip.de TAP query.
• Full schemas re-validated 2026-07-20 against TAP_SCHEMA + the SHboost Parquet

footer → references/schema.md (SHboost column list corrected: xgb_* naming, no *50 columns).

• Data-paper citations + accountable curator/ORCID added 2026-07-21 (Skill Commons

first-package publication prep).

Source & license

This open-source skill is cataloged on AgentStack and links to its original source — we do not rehost the code.

• Author: skill-commons
• Source: skill-commons/skill-commons
• License: Apache-2.0

Install and usage instructions live in the source repository linked above.