astro-legacy-archive/planck-2018-chains

Planck 2018 Cosmological-Parameter Chains

This dataset mirrors the Planck Public Release 3 cosmological-parameter full grid, COM_CosmoParams_fullGrid_R3.01.zip. It contains the Markov chains and their GetDist and CosmoMC companions for 336 combinations of cosmological model and likelihood or external-data selection. The 1,296 chain roots become 5,184 Parquet tables containing 27,699,519 rows.

The tables retain the source's headerless, positional structure. Arrow fields are named Column 1, Column 2, and so on; parameter names from the source .paramnames companion are explanatory metadata and are not substituted for those field identities.

Release	Planck PR3 / 2018 cosmological-parameter full grid
Configurations	336 model/likelihood directory pairs
Chain roots	1,296
Parquet tables	5,184; four numbered chains per root
Rows	27,699,519
Positional width	18--122 float64 columns, depending on root
Source archive	COM_CosmoParams_fullGrid_R3.01.zip

Finding a chain

A configuration is one source model directory and one source likelihood-combination directory joined by exactly two underscores:

<model_dir>__<likelihood_dir>

For example, the source directory base/plikHM_TTTEEE_lowl_lowE/ is configuration base__plikHM_TTTEEE_lowl_lowE. The left side selects the cosmological model. The right side retains the release's likelihood spelling, which can include Planck temperature, polarization, low-multipole, lensing, BAO, supernova, distance-ladder, DES, or BK15 components. Neither side is parsed or normalized by this mirror.

Each configuration directory contains one or more chain roots. The root matching the example above is base_plikHM_TTTEEE_lowl_lowE. Processed roots such as base_plikHM_TTTEEE_lowl_lowE_post_BAO are separate source identities in the same configuration. A root always has these four tables:

base_plikHM_TTTEEE_lowl_lowE_1.txt.parquet
base_plikHM_TTTEEE_lowl_lowE_2.txt.parquet
base_plikHM_TTTEEE_lowl_lowE_3.txt.parquet
base_plikHM_TTTEEE_lowl_lowE_4.txt.parquet

The complete artifact path is therefore <config>/<verbatim chain filename>.parquet. Start with a model and data combination, enter its double-underscore configuration, choose the exact root, then select one or all four numbered tables.

The directory names are the grid's index. Begin with base for the six-parameter model, or choose a source model spelling such as base_omegak, base_mnu, base_r, base_w, or base_w_wa when that extension is the target. Then choose the likelihood directory whose components match the intended data combination. Configurations contain between one and 16 roots. A plain root is the corresponding chain exploration; _post_* roots are the release's separately stored processed chains and remain separate here. For example, base__plikHM_TTTEEE_lowl_lowE contains the plain root together with the source's _post_BAO, _post_Riess18, _post_zre6p5, and combined processed variants. Consult the exact filenames rather than assuming that every model or likelihood combination has the same root set.

Select one root at a time when using load_dataset. Different roots inside a configuration can have different source widths; the example configuration spans 94--99 columns. Its roots are distinct tables, so combining every Parquet in the directory into one inferred schema is not a valid representation. Supplying one root's four filenames keeps the source schema and combines only its four chain segments.

How to use

python -m venv .venv && .venv/bin/pip install datasets huggingface_hub pyarrow

from datasets import load_dataset
root = "base_plikHM_TTTEEE_lowl_lowE"
files = [f"base__plikHM_TTTEEE_lowl_lowE/{root}_{n}.txt.parquet" for n in range(1, 5)]
chain = load_dataset("astro-legacy-archive/planck-2018-chains", "base__plikHM_TTTEEE_lowl_lowE", data_files=files, split="train[:5]")
print(chain[0]["Column 30"])

66.62354

import json
from huggingface_hub import hf_hub_download
import pyarrow.parquet as pq
path = hf_hub_download("astro-legacy-archive/planck-2018-chains", "base__plikHM_TTTEEE_lowl_lowE/base_plikHM_TTTEEE_lowl_lowE_1.txt.parquet", repo_type="dataset")
print(json.loads(pq.read_schema(path).metadata[b"statistical_product"])["schema"])

astro-legacy-archive/statistical-product

Reading the positional columns

The embedded format note pins the CosmoMC convention weight like param1 param2 param3 ...: Column 1 is the source weight, Column 2 is minus the log likelihood, and the .paramnames rows correspond in order to Column 3 onward. Resolve a parameter by finding its exact row in the companion and adding two to its one-based row number. Preserve suffixes such as the source's H0* and omegam*; in the example root they describe Column 30 and Column 32. This association explains a position without changing the stored field name.

Every Parquet embeds an ordered self_description.companions list. It includes the root's verbatim .paramnames, .ranges, .properties.ini, GetDist .margestats, .likestats, .converge, and .covmat, plus the applicable best-fit, likelihood, input, theory, and foreground text products. Each record carries its exact relative path, UTF-8 content, byte count, and SHA-256. The source ZIP pin and the checked CosmoMC format document are embedded alongside them.

The preview uses all 24,497 rows from the four unprocessed base_plikHM_TTTEEE_lowl_lowE tables. Its axes are resolved from the verbatim H0* and omegam* rows; marker area reflects the source log(1 + weight) and colour shows the source negative log likelihood.

Source fidelity and provenance

The full-grid ZIP has SHA-256 6d4775fa7b59a9387f630fc0187d463c7b95c4d97d7b9e317eebb0744b2962fe. Every output was rebuilt in a different directory and compared by relative path, size, and SHA-256. A separate source-backed traversal then checked all 5,184 tables against the ZIP, including row count, positional width, Arrow names, companion order and content, and float64 value bytes. The retained tables total 22,970,406,175 bytes; their ordered artifact digest is 4954825b1e4966d6a9918fec37c288b0dc207521f93c84e7984ec5b1f72439e3.

The release contains 72 repeated paths under four chain roots. Both central-directory occurrences are retained in self-description; the 16 repeated numbered-chain paths were represented once only after their occurrence bytes compared equal. The sole source .paramnames~ editor backup and the plik, camspec, and BKPLANCK foreground companion spellings are also retained.

Licensing, credit, and citation

ESA states:

Data hosted in the ESA Space Science Archives are distributed under the CC BY-NC 3.0 IGO license.

ESA's Planck publications page requests:

If you use any of these results for presentations, please acknowledge the corresponding paper, ESA/Planck, and the Planck Collaboration.

Please cite Planck Collaboration, "Planck 2018 results. VI. Cosmological parameters", Astronomy & Astrophysics 641, A6 (2020), arXiv:1807.06209. The ESA dataset record is PR3 Cosmology Products. The archived Parquet release is 10.57967/hf/10268.

Errata and requests

To report an erratum, propose a correction, or ask a question about this dataset, open a Discussion or Pull Request in its Community tab; posting requires a free Hugging Face account. Dataset requests and archive-wide questions may be posted in the README space's Community tab or the Community tab of any dataset. Confirmed corrections are made in the source conversion tools and published by rebuilding the dataset.