| --- |
| language: |
| - zh |
| - en |
| pretty_name: MATTER-604 Evaluator Access |
| license: apache-2.0 |
| size_categories: |
| - n<1K |
| task_categories: |
| - question-answering |
| tags: |
| - materials-science |
| - agent |
| - benchmark |
| - private-evaluation |
| gated: manual |
| extra_gated_heading: Request access to the MATTER-604 self-hosted evaluator |
| extra_gated_description: Access is intended for trusted benchmark maintainers who will isolate hidden answers and rubrics from evaluated Agents. |
| extra_gated_prompt: Approved users receive a downloadable evaluator snapshot. Results must be reported as self-hosted rather than official MATTER scores. |
| extra_gated_fields: |
| Full name: text |
| Affiliation: text |
| Intended evaluation use: text |
| Agent or model to be evaluated: text |
| Answer isolation and anti-cheating plan: text |
| Planned publication of results or traces: text |
| I understand that the evaluated Agent must not access hidden answers, rubrics, or evaluator internals: checkbox |
| I understand that results produced with this package are self-hosted and not official MATTER scores: checkbox |
| --- |
| |
| # MATTER-604 Evaluator Access |
|
|
| This manual-gated repository provides a versioned copy of the MATTER-604 |
| reference answers, per-task rubrics, Gold fixtures, and evaluator code to |
| trusted third-party benchmark maintainers. |
|
|
| Scores produced from this package are **self-hosted MATTER evaluations**, not |
| official scores issued by the canonical private evaluator maintained by |
| MatMaster-DP. See `ACCESS_POLICY.md` for the evidence and reporting contract. |
|
|
| ## Repository contents |
|
|
| ```text |
| MATTER-604-evaluator-access/ |
| ├── README.md |
| ├── ACCESS_POLICY.md |
| ├── LICENSE |
| ├── manifest.json |
| ├── requirements.txt |
| ├── restricted_assets.json |
| ├── score-config.yaml |
| ├── tasks.tsv |
| ├── evaluation/ |
| │ ├── __init__.py |
| │ ├── config.yaml |
| │ ├── core/ |
| │ ├── question_bank/ |
| │ ├── scripts/ |
| │ └── validators/ |
| └── .gitattributes |
| ``` |
|
|
| | Path | Description | |
| |---|---| |
| | `README.md` | Dataset Card describing the evaluator package, repository layout, scoring components, and access workflow. | |
| | `ACCESS_POLICY.md` | Operating and reporting contract for answer isolation, evidence retention, self-hosted evaluation, and Bohr CLI traces. | |
| | `LICENSE` | Apache-2.0 license for MATTER-owned content. | |
| | `manifest.json` | Machine-readable package identity: 604 tasks, manual-gated release type, evaluator commitment, scoring authority, and resource metadata. | |
| | `requirements.txt` | Pinned Python dependencies required by the evaluator and scientific validators. | |
| | `restricted_assets.json` | Identity, size, SHA-256, provenance, and acquisition records for six license-controlled scientific inputs. | |
| | `score-config.yaml` | Scoring configuration with equal correctness, grounding, and efficiency weights plus the evaluator-LLM runtime settings. | |
| | `tasks.tsv` | Ordered index of 604 unique task IDs used to select and align question-bank records. | |
| | `evaluation/` | Hidden question bank, reference answers, rubrics, evidence model, scoring engine, validators, and scoring scripts. | |
| | `.gitattributes` | Hugging Face repository transport metadata managed by the Hub. | |
|
|
| ## Evaluation package |
|
|
| The `evaluation/` tree contains 406 repository files organized as follows: |
|
|
| | Component | Files | Description | |
| |---|---:|---| |
| | `evaluation/core/` | 15 | Evaluator construction, schemas, evidence records, scoring orchestration, LLM utilities, task loading, and score aggregation. | |
| | `evaluation/question_bank/` | 363 | Sixty question-bank YAML files containing 604 unique tasks, their reference-answer fields, 5,288 checklist criteria, and the scientific inputs used by the evaluator. | |
| | `evaluation/scripts/` | 3 | Entry points for devshell scoring, external-baseline scoring, and audited Bohr CLI receipt collection. | |
| | `evaluation/validators/` | 23 | Deterministic checks for structures, compositions, densities, distances, planarity, VASP, ABACUS, GPUMD, GROMACS, DP-GEN, JSON/text artifacts, Bohr CLI operations, and run budgets. | |
| | `evaluation/config.yaml` | 1 | Evaluator defaults and runtime configuration consumed by the scoring engine. | |
| | `evaluation/__init__.py` | 1 | Python package entry point for the evaluation modules. | |
|
|
| All 604 tasks contain a scoring checklist. The question bank contains 5,288 |
| criteria across correctness, grounding, and efficiency. Reference-answer fields |
| are populated for 598 tasks; the six safety-refusal tasks encode their expected |
| behaviour through their checklists. |
|
|
| ## Scoring scripts |
|
|
| ```text |
| evaluation/scripts/baseline/score_baseline_tasks.py |
| evaluation/scripts/devshell/score_devshell_tasks.py |
| evaluation/scripts/devshell/bohr_cli_audit.py |
| ``` |
|
|
| - `score_baseline_tasks.py` builds evidence from an external Agent workspace and |
| retained event stream, then evaluates the task criterion by criterion. |
| - `score_devshell_tasks.py` reads devshell workspaces, summaries, events, and |
| process receipts before producing axis scores and per-criterion reasons. |
| - `bohr_cli_audit.py` wraps the real `bohr` executable during solving and writes |
| structured `bohr_cli_receipts.jsonl` evidence for platform-operation checks. |
|
|
| ## Scientific validators |
|
|
| The validator collection covers CIF/POSCAR/STRU parsing, composition, density, |
| atomic distances, ordering, molecular crystals, planarity, VASP INCAR, ABACUS |
| inputs, GPUMD `run.in`, GROMACS topologies, DP-GEN arguments, KPT files, JSON and |
| text outputs, Bohr CLI receipts, artifact budgets, token budgets, and turn budgets. |
|
|
| ## Runtime dependencies |
|
|
| The evaluator pins ASE, pymatgen, NumPy, Pydantic, PyYAML, JSON Schema, DP-GEN, |
| Molcrys-kit, OpenAI, and Requests in `requirements.txt`. `score-config.yaml` |
| uses equal weights for correctness, grounding, and efficiency and configures the |
| evaluator LLM as `deepseek-v4-pro` with deterministic temperature `0.0`. |
|
|
| ## Bohr CLI tasks |
|
|
| Fifty-three of the 604 tasks are tagged `bohr-cli`. Their solve environment |
| must contain an authenticated real Bohr CLI and the evaluation audit launcher |
| must preserve `bohr_cli_receipts.jsonl`. The published three-Agent runs used |
| Node.js 22 and `@dptech-corp/bohr-cli@2.5.17`. The devshell scorer reads the |
| retained receipts as part of the self-hosted evidence package. |
|
|
| ## License |
|
|
| MATTER-owned content in this repository is provided under Apache-2.0. |
|
|
