README.md

metadata

language:
  - zh
  - en
pretty_name: MATTER-604 Public Questions
license: apache-2.0
size_categories:
  - n<1K
task_categories:
  - question-answering
tags:
  - materials-science
  - agent
  - benchmark
  - hidden-test
configs:
  - config_name: default
    data_files:
      - split: test
        path: data/test.jsonl

MATTER-604 Public Questions

MATTER-604 evaluates materials-science Agents on structure construction and retrieval, scientific analysis, workflow orchestration, input generation, data diagnosis, execution contracts, batch processing, and safety refusal.

This repository contains the 604 public task prompts and declared input fixtures. It intentionally does not contain per-task reference answers, checklists, Gold thresholds, or the private evaluator. Submit answers and task artifacts through the benchmark evaluation service.

Repository contents

MATTER-604/
├── README.md
├── EVALUATION_GUIDE.md
├── LICENSE
├── data/
│   └── test.jsonl
├── fixtures/
├── manifest.json
├── private_eval_commitment.json
├── submission.schema.json
└── .gitattributes

Path	Description
README.md	Dataset Card: scientific scope, task distribution, runtime requirements, scoring summary, and release boundaries.
EVALUATION_GUIDE.md	Provider-neutral instructions for preparing one-task workspaces, invoking an Agent, recording traces and artifacts, and producing schema-valid submissions.
LICENSE	Apache-2.0 license for MATTER-owned content. It does not license excluded VASP POTCAR files or third-party PDFs.
data/test.jsonl	The 604 evaluation tasks, one JSON object per line. Each record contains only task_id, capability, domain, prompt, tags, and data_files.
fixtures/	238 redistributable task inputs declared by 147 tasks, including crystal structures, simulation inputs, spectra, electrochemical data, tabular data, and small model/test files.
manifest.json	Machine-readable release identity: task count, task-ID hash, fixture counts, hidden fields, missing-input status, license, and closed-source-runtime exclusion.
private_eval_commitment.json	A whole-tree SHA-256 commitment to the canonical 604-task private evaluator. It identifies the scoring version without exposing answers or file-level hashes.
submission.schema.json	JSON Schema for third-party submissions: final answer, artifact paths and hashes, optional trace path, and Agent/model/prompt metadata.
.gitattributes	Hugging Face repository transport metadata added by the Hub; it is not part of the scientific benchmark content.

The 147 tasks with declared inputs contain 254 file references to 244 unique logical paths. Of these, 238 files are bundled under fixtures/; six restricted inputs are represented only by their identity, SHA-256, size, restriction reason, source, and lawful acquisition instructions.

To run an Agent on the benchmark, follow EVALUATION_GUIDE.md.

Capability distribution

The capability label describes the main research action being evaluated. A task can exercise supporting skills in addition to its primary label.

Capability	Tasks	Share	What is evaluated
input_generation	164	27.2%	Producing scientifically consistent inputs for codes such as VASP, ABACUS, CP2K, LAMMPS, GROMACS, ORCA, DP-GEN, and GPUMD.
workflow_orchestration	138	22.8%	Organizing multi-step modelling or calculation workflows, remote jobs, state transitions, recovery, and result collection.
scientific_analysis	95	15.7%	Extracting and interpreting materials properties from structures, energies, spectra, trajectories, electrochemical data, and other scientific outputs.
structure_construction	84	13.9%	Building chemically and crystallographically meaningful bulk, molecular, surface, interface, defect, and other atomistic models.
structure_retrieval	49	8.1%	Finding, selecting, converting, and validating material or molecular structures from databases and supplied sources.
execution_contract	46	7.6%	Respecting requested methods and constraints, checking feasibility, handling infrastructure limitations, and avoiding silent substitutions or unsupported claims.
data_diagnosis	18	3.0%	Identifying malformed inputs, inconsistent metadata, convergence failures, corrupted outputs, and scientifically invalid intermediate states.
safety_refusal	6	1.0%	Refusing unsafe, unauthorized, deceptive, or scientifically unjustified operations while explaining the boundary.
batch_processing	4	0.7%	Managing related task sets, controlled parameter sweeps, consistent batch inputs, and batch-level failure recovery.
Total	604	100%

Domain distribution

The domain label identifies the main materials context. agnostic covers cross-domain scientific-software and research-operation tasks rather than a separate material class.

Domain	Tasks	Share	Scope
agnostic	467	77.3%	General atomistic modelling, electronic-structure, molecular simulation, data analysis, scientific software, and platform operations.
battery	59	9.8%	Electrodes, electrolytes, ion transport, cycling data, interphases, and battery-oriented simulations.
semiconductor	27	4.5%	Electronic structure, dielectric response, defects, band properties, and semiconductor materials.
catalysis	19	3.1%	Surfaces, adsorption, reaction energetics, electrocatalysis, and catalytic screening.
polymer	19	3.1%	Polymer and molecular materials, force fields, conformations, membranes, and polymer simulations.
alloy	13	2.2%	Alloy structures, ordering, solid solutions, magnetic states, and alloy-property workflows.
Total	604	100%

Splits

data/test.jsonl contains the 604 evaluation questions. The test labels are private. fixtures/ contains only local files explicitly declared by a task's data_files field.

Six declared inputs are not copied into this repository: two VASP POTCAR files and four third-party publisher PDFs. Their records use distribution=restricted and provide the exact SHA-256 plus lawful acquisition instructions. A runner must provision those files outside the dataset repo and verify their hashes before starting the affected task.

Scoring method

Each private checklist item receives a binary pass/fail verdict and an optional weight. Within each active axis:

axis_score = sum(pass_i * item_weight_i) / sum(item_weight_i)

The overall score is the weighted mean of active correctness, grounding, and efficiency axes. Exact per-task criteria, expected values, tolerances, and Gold files remain private to prevent direct answer lookup.

Intended use and contamination policy

Use this dataset for evaluation, not training or answer distillation. Public availability cannot technically prevent training; benchmark users should disclose any exposure to MATTER questions or model answers. The official score is produced only by the versioned private evaluator identified in private_eval_commitment.json.

Runtime

The closed-source MatMaster image is not distributed. Third-party Agents may use their own runtime as long as they receive only the public task package and return the documented answer/artifact bundle.

Bohr CLI

Of the 604 tasks, 53 are tagged bohr-cli. A complete run must provide an authenticated real bohr executable for those tasks and retain the audited bohr_cli_receipts.jsonl produced during solving. The other 551 tasks do not have a benchmark-wide mandatory Bohr CLI dependency, although Bohrium can be used as an optional remote-compute route when the task contract permits it.

The published three-Agent runs used Node.js 22 and Bohr CLI 2.5.17. Reproduce that environment with:

npm install -g @dptech-corp/bohr-cli@2.5.17
bohr version -o json

An unpinned npm install -g @dptech-corp/bohr-cli follows npm latest and must be reported as a different runtime if its version differs. The devshell scoring path does not call Bohrium again; it consumes receipts and retained tool events from the solve. Installing the CLI after a run cannot reconstruct missing execution evidence. The included external baseline scorer does not by itself load the process-level receipt file or reproduce the same global Bohr execution gate, so it is not a drop-in equivalent for the 53-task Bohr slice. A third-party adapter must preserve and ingest equivalent receipt evidence. Keep the authentication cache and access key outside Agent outputs.

License

MATTER-owned benchmark content in this repository is released under Apache-2.0. Restricted VASP POTCAR files and third-party publisher PDFs are not included and are not covered by this repository's Apache-2.0 license.