astromindinc/axia-csc-corpus

Axia — Chandra Source Catalog corpus

astromindinc/axia-csc-corpus — 51,450 X-ray sources from the Chandra Source Catalog 2.1, each carrying:

• Per-photon event lists in two forms: an event_list pruned to a single 8 h window in 0.5-8 keV (the input shape the Axia fine-tuned model trained on), and an original_event_list containing the full unpruned observation (the input to the model-free spectrum-snapshot / light-curve pipeline).
• A 64-d learned embedding (pca_64d) suitable for nearest-neighbour / vector search. The values are cosine-similarity-ready.
• A 2-d UMAP projection (umap_2d) for visualisation.
• The full CSC catalog metadata: hardness ratios, four-model spectral fits, photometry, variability indices, sky coordinates, etc.

This is the dataset that backs the Axia multi-agent X-ray source decoder.

Quick start

from datasets import load_dataset

ds = load_dataset("astromindinc/axia-csc-corpus", split="train")
print(ds)
# Dataset({
#     features: [...],
#     num_rows: 51,450
# })

src = ds[0]
print(src["source_name"], "@", (src["ra"], src["dec"]))
print(" pruned events:", len(src["event_list"]))
print(" full events:  ", len(src["original_event_list"]))
print(" embedding:    ", len(src["pca_64d"]), "dim")

Stratification

The 11 source-type categories present in the corpus (training-set distribution):

Category	Count
Other	15,849
Large accretors	12,604
Young stars	8,221
Normal stars	4,407
Small accretors	3,899
Normal galaxies	2,549
Active / variable stars	1,386
Stellar systems & clusters	1,207
Stellar remnants	555
Massive stars	461
White Dwarf accretors	312

Fields

Field	Type	Notes
obsid	int	Chandra observation ID
obi	int	Observation interval
region_id	int	CSC region ID
source_name	string	CSC source designation (e.g. 2CXO J123456.7+001122)
ra, dec	float	ICRS / J2000, decimal degrees
theta	float	Off-axis angle, arcmin
source_type	string	CSC source_type (e.g. AGN, Seyfert1, XB, X, ...)
source_type_category	string	One of 11 broad categories used for stratification
thermal_classification	string	thermal / nonthermal
event_list	list[[t_s, energy_eV]]	Pruned event list — single 8 h window, filtered to 0.5–8 keV. This is what the Axia fine-tuned model trained on.
original_event_list	list[[t_s, energy_eV]]	Original event list — full observation, all energies. Used by the spectrum-snapshot / light-curve / dE-dt computations.
pca_64d	listfloat	64-d embedding from the Axia XrayProcessor → PCA pipeline. Cosine-similarity-indexable.
umap_2d	listfloat	2-d UMAP projection for visualisation
hard_hs, hard_hm, hard_ms	float	Hardness ratios in the standard CSC H/M/S bands
flux_significance_b	float	Broad-band detection significance
flux_aper_b, flux_bb_aper_b	float	Broad-band aperture fluxes
src_cnts_aper_b	float	Net source counts in the aperture
var_index_b, var_prob_b	float	Variability index / probability (CSC)
gti_mjd_obs	float	Observation start time in MJD
powlaw_stat, powlaw_gamma, powlaw_nh, powlaw_ampl	float	Power-law spectral fit
powlaw_gamma_lolim, powlaw_gamma_hilim	float	Power-law Γ confidence bounds
powerlaw_gamma_low, powerlaw_gamma_high	float	Alternate Γ bound spelling (kept verbatim from CSC)
bb_stat, bb_kt, bb_nh, bb_ampl	float	Black-body spectral fit
brems_stat, brems_kt	float	Bremsstrahlung spectral fit
apec_stat, apec_kt, apec_nh, apec_norm, apec_abund, apec_z	float	APEC plasma spectral fit (NaN→null when not fit)
preferred_spectral_model	list[string]	Catalog recommendation
recommended_model	string	Catalog recommendation
match_type	string	CSC master-source match type
significance	float	CSC source significance

Floats that are NaN in the underlying CSC catalog are stored as JSON null. ObjectIds and datetimes have been converted to strings.

Provenance

The X-ray and catalog data are derived from the Chandra Source Catalog 2.1 (CSC 2.1), which is freely available from the Harvard CXC. The pca_64d embedding and the umap_2d projection were computed using the Axia fine-tuned model (DeepSeek-R1-Distill-Qwen-7B + XrayProcessor, LoRA r=8).

Source cluster of this dump: am-webapp.niojd.mongodb.net. Produced: 2026-05-28T10:31:23.830768Z.

Auxiliary files in this repo

• data/corpus.jsonl.gz — the main file. One JSON document per line.
• data/metadata_records.json — small dataset registry used by the Axia webapp's /api/datasets endpoint.
• data/atlas_indexes/pca_64_vector_search.json — MongoDB Atlas Vector Search index definition for the pca_64d field (cosine, 64 dims).
• manifest.json — original dump provenance (counts, SHA-256, source cluster identifier).

Loading into MongoDB

The companion script data/ingest/load_from_huggingface.py in the Axia repo downloads this dataset and loads it into a MongoDB instance (local or Atlas), creating the vector-search index if the target is Atlas.

# In the axia/ repo:
make load-from-hf                       # default: astromindinc/axia-csc-corpus
make load-from-hf DATASET=...           # custom HF repo id

License

This packaged dataset is released under CC-BY-4.0. The underlying Chandra Source Catalog data is in the public domain; see the CSC homepage for the original data use policy. Please cite both the Axia paper and the CSC if you use this dataset in published work.

Citation

@misc{axia2026,
  title     = {Axia: a multi-agent decoder for Chandra X-ray sources},
  author    = {AstroMind Authors},
  year      = {2026},
  note      = {TBA — paper in preparation}
}