README.md

metadata

pretty_name: WxFlowBench
language:
  - en
  - zh
tags:
  - Weather
  - Agent
  - Operation

WxFlowBench pre-release snapshot (2026-09-24)

This repository preserves the latest locally materialized pre-release results for tracks T0–T4. It is a data snapshot, not a final benchmark release. The task questions, hidden split, evaluator answers, and scoring materials are all public here by the dataset owner's decision. Do not use this snapshot as a blind hidden-test benchmark.

Layout

Folder	Snapshot	Source files	Source bytes	Status
T0/	run006, 552 cases (Core512, RAD20, SAT20), plus the locally bound climatology shards and RAD/SAT indexes	1,198	3,179,325,467	Pre-release; human adjudication and full trajectory/baseline validation remain
T1/	U.S. T1/T2 release package v4 and its exact payload inventory	8,585	1,186,367,350	170 T1 tasks; production admission not granted
T2/	Same source package as T1, retained under this track for self-contained access	8,585	1,186,367,350	192 T2 tasks; production admission not granted
T3/	France and U.S. cumulative pre-release lineages, through spring_integrated_prerelease_v1 and iowa_freeze_us_prerelease_v1	8,478	114,806,993,258	Internal pre-release; source/evaluation/release readiness false
T4/	Same combined T3/T4 lineage, retained under this track for self-contained access	8,478	114,806,993,258	Internal pre-release; source/evaluation/release readiness false

Shared packages are duplicated under T1/T2 and T3/T4 because their original source contracts combine both tracks. Each track directory contains manifest.json and uncompressed .tar archives. Paths inside the archives retain their original repository-relative names. The manifest records every source path and each archive's SHA-256 digest. Packet archives target roughly 650 MB of source files. Dependency archives target roughly 1 GB where the source file size allows; the largest contains one 19.8 GB source file. T3/T4 packet archives are grouped by their package files' UTC modification dates; referenced source files are in separate dependency-part-*.tar archives.

To retrieve a track:

hf download when665/WxFlowBench --repo-type dataset --include 'T0/*' --local-dir ./WxFlowBench
cd WxFlowBench
sha256sum -c T0/SHA256SUMS
for archive in T0/*.tar; do tar -xf "$archive"; done

The T1/ and T2/ manifests refer to the same U.S. package; the T3/ and T4/ manifests refer to the same combined lineage. You only need one copy of each pair when extracting all five tracks. The source paths in each archive are relative to the original WeatherAgent repository root.

Scope and known limits

• T0 run006 contains 281 public and 271 hidden Agent files, 552 evaluator files, and one inventory. Its source-bound climatology shards and RAD20 parent/coordinate data are included. The broader weather source catalogs and a standalone execution environment are not bundled.
• T1/T2 v4 contains release.json, payload-inventory.json, and all 8,583 listed payload files. It covers U.S. tasks only; France is not in this release package. Six tasks have incomplete individual sources as recorded in release.json.
• T3/T4 includes the two latest cumulative France and U.S. chains plus their locally materialized JSON path dependencies and in-directory support files. The France tip reports 69 T3 and 78 T4 technically accepted tasks; the U.S. tip reports 25 T3 and 22 T4 proposed/accepted task counts. These are construction counts, not a final evaluated test set. All current packages mark source_ready, evaluable, and release_ready false and expert review pending. References to sources that were not materialized locally remain external; the manifest lists exactly what is present in this snapshot.
• MeteoNet material is credited to MÉTÉO FRANCE: original data from https://meteonet.umr-cnrm.fr/, last source update 30 January 2020; MeteoNet is offered under Etalab Open Licence 2.0 (https://meteonet.umr-cnrm.fr/assets/open-licence.pdf). Please cite Larvor et al., MeteoNet, an open reference weather dataset by METEO FRANCE (2020). Other third-party source data and reports retain their respective upstream terms. No single license is asserted for the full collection.

The original JSON files and their internal hashes are preserved byte for byte. The per-track manifests and archive checksums describe this upload snapshot; they do not change the readiness flags or grant production admission.