Rewrite public dataset card

Document provenance, structure, seed overrides, public reference-answer leakage, intended use, limitations, download instructions, and the absence of a dataset-wide upstream license.

README.md

@@ -1,26 +1,230 @@
 ---
-pretty_name: ASI-Bench Instances (seed 31415 with seed 2 overrides)
+pretty_name: ASI-Bench Generated Instances (Seed 31415 Snapshot)
+language:
+  - en
+license: other
+license_name: no-dataset-wide-license-declared
+size_categories:
+  - n<1K
+annotations_creators:
+  - machine-generated
+source_datasets:
+  - original
+tags:
+  - ai-for-science
+  - benchmark
+  - scientific-reasoning
+  - agents
+  - agent-evaluation
+  - code-generation
+  - multimodal
+  - text
+  - tabular
+  - timeseries
 ---
 
-# ASI-Bench instances — seed 31415 with seed 2 overrides
+# ASI-Bench Generated Instances
 
-This dataset contains the original seed 31415 instances, with seed 2 instances for
-`computer_science.deployment_prediction_sets` and
-`math.ucb_q_learning_regret`. The latter is an explicitly requested `test` task,
-so the dataset contains 60 tasks after the requested removal of
-`computer_science.muon_optimizer_replay_audit`. Each instance includes four rendered
-prompts, public input data, and the private reference answer used for local
-validation. Exact seed overrides are recorded in `metadata.json`.
+This repository is a file-based snapshot of generated task instances from
+[Agent-AI4Sci-Bench](https://github.com/apexin-ai/Agent-AI4Sci-Bench), a
+benchmark for evaluating agents on project-level AI-for-Science problems.
+
+The snapshot contains **60 task instances**, **1,530 files**, and approximately
+**113 MiB** of data. Each task provides four prompt variants, public input data,
+and generated reference artifacts. It is intended for benchmark development,
+reproducibility studies, scorer validation, and local evaluation research.
+
+> **Answer-leakage warning:** every task directory contains `reference/`
+> artifacts. These are ground-truth or validation materials and are public in
+> this repository. Do not expose them to an evaluated agent, and do not use this
+> snapshot for a blind public leaderboard. Use separately generated private
+> instances for contamination-resistant evaluation.
+
+## Snapshot provenance
+
+| Field | Value |
+|---|---|
+| Source repository | `apexin-ai/Agent-AI4Sci-Bench` |
+| Source commit | `6192234d4b9fdd8657a8054ac12b4cbaa827ea26` |
+| Number of tasks | 60 |
+| Source status | 59 `final` tasks and 1 explicitly requested `test` task |
+| Default generation seed | 31415 |
+| Reference answers included | Yes |
+| Generated | 2026-08-03 (UTC) |
+
+`metadata.json` is the authoritative machine-readable manifest for task IDs,
+source provenance, and generation seeds.
+
+### Seed exceptions
+
+Two instances were generated with **seed 2**, while their directory names
+retain the dataset-wide `__seed31415` suffix:
+
+- `computer_science.deployment_prediction_sets__seed31415`
+- `math.ucb_q_learning_regret__seed31415`
+
+Their actual seed is recorded in `metadata.json` under `seed_overrides`. Folder
+suffixes must not be used as the sole source of generation provenance.
+
+`math.ucb_q_learning_regret` was marked `test`, rather than `final`, in the
+recorded source snapshot. `computer_science.muon_optimizer_replay_audit` is not
+included in this published dataset.
+
+## Scientific domains
+
+| Domain | Tasks |
+|---|---:|
+| Astronomy | 3 |
+| Biology | 6 |
+| Biostatistics | 1 |
+| Chemistry | 6 |
+| Computer science | 9 |
+| Earth science | 4 |
+| Electrical engineering | 2 |
+| Materials science | 2 |
+| Mathematics | 12 |
+| Physics | 8 |
+| Robotics | 7 |
+| **Total** | **60** |
+
+## Repository structure
+
+This is not a conventional row-oriented dataset and does not define standard
+train, validation, or test splits. Each top-level task directory is one
+generated benchmark instance:
 
 ```text
-<task_id>__seed31415/
+<domain>.<task_id>__seed31415/
 ├── prompt_b1.md
 ├── prompt_b2.md
 ├── prompt_b3.md
 ├── prompt_b4.md
 ├── data/
-└── reference/
+├── reference/
+├── instance_meta.json
+└── framework_task_info.json
+```
+
+- `prompt_b1.md` through `prompt_b4.md` are four disclosure levels of the task
+  prompt.
+- `data/` contains the files that may be provided to the evaluated agent.
+- `reference/` contains generated answers, diagnostics, or other validation
+  artifacts and must be isolated from the agent during evaluation.
+- `instance_meta.json` and `framework_task_info.json` record instance and
+  framework provenance.
+
+Schemas vary by task. Files include Markdown, JSON, CSV, Python, NumPy arrays
+(`.npy` and `.npz`), and images. Consumers should treat each task as a directory
+contract rather than concatenate the repository into a single table.
+
+## Download
+
+Using the Hugging Face CLI:
+
+```bash
+hf download Apexintelligence-AI/ASI-Bench-seed31415 \
+  --repo-type dataset \
+  --local-dir ASI-Bench-seed31415
 ```
 
-The two overridden task directories retain the dataset suffix `__seed31415`;
-their actual generation seed remains recorded in `metadata.json`.
+Or with Git and Git LFS:
+
+```bash
+git lfs install
+git clone https://huggingface.co/datasets/Apexintelligence-AI/ASI-Bench-seed31415
+```
+
+## Recommended evaluation use
+
+1. Select exactly one prompt level (`b1`–`b4`) for a run.
+2. Give the agent only that prompt and the corresponding `data/` directory.
+3. Keep `reference/` and framework-only metadata outside the agent's readable
+   workspace.
+4. Run the matching scorer from the recorded Agent-AI4Sci-Bench source commit.
+5. Record the task ID, prompt level, source commit, actual generation seed,
+   runtime, dependencies, and agent configuration with every result.
+
+The repository contains instances and reference artifacts, not a complete
+leaderboard protocol. Reproducible comparisons additionally require the
+matching benchmark code, scorer configuration, runtime dependencies, and
+execution limits.
+
+## Intended uses
+
+- Reproducing and auditing AI-for-Science benchmark tasks.
+- Developing agent harnesses, sandboxes, and scorer integrations.
+- Studying how prompt disclosure level affects agent behavior.
+- Testing data loading and output validation for heterogeneous scientific
+  workflows.
+- Creating demonstrations or tutorials where public reference answers are
+  acceptable.
+
+## Out-of-scope uses
+
+- Reporting this public snapshot as a contamination-resistant benchmark.
+- Allowing an evaluated agent to read `reference/` or framework-only metadata.
+- Treating scores from different prompt levels, scorer versions, or runtime
+  limits as directly comparable without disclosure.
+- Assuming the task collection represents all scientific fields, methods, or
+  real-world deployment conditions.
+
+## Limitations and responsible use
+
+- The snapshot contains only one generated instance per task, so it does not
+  measure variation across seeds.
+- Two tasks use a generation seed different from their folder suffix; consult
+  `metadata.json`.
+- One included task has source status `test`, while the others have status
+  `final`.
+- Public reference artifacts create a substantial risk of evaluation leakage,
+  memorization, and benchmark gaming.
+- Tasks have heterogeneous schemas, dependencies, resource demands, scoring
+  logic, and numerical tolerances.
+- Prompts are in English and the domain distribution is not balanced.
+- Generated or simulated scientific tasks may simplify assumptions found in
+  real research workflows and should not be used to support scientific,
+  medical, safety-critical, or policy decisions without independent review.
+- The dataset was not designed to contain personal information, but no formal
+  privacy audit is claimed. Users publishing derived artifacts should review
+  them for unintended sensitive content.
+
+## Licensing
+
+The recorded upstream source commit does **not** declare a repository-wide
+license. This dataset therefore uses the Hugging Face `other` license label and
+does not grant a blanket license for reuse, redistribution, model training, or
+commercial use. Public availability alone should not be interpreted as
+permission.
+
+Some task inputs, dependencies, or bundled components may have their own terms.
+Users are responsible for reviewing the
+[source repository](https://github.com/apexin-ai/Agent-AI4Sci-Bench), relevant
+task directories, third-party notices, and applicable law before reuse. Dataset
+maintainers should add an explicit repository-level license before public
+release if broader permissions are intended.
+
+## Maintenance and versioning
+
+This is a pinned generated snapshot, not a continuously synchronized mirror of
+the upstream `main` branch. Changes to the upstream benchmark after the recorded
+commit are not automatically reflected here. Consult `metadata.json` and the
+Hugging Face commit ID when reporting results.
+
+For questions or corrections, use the repository's Hugging Face Discussions or
+the upstream GitHub issue tracker.
+
+## Citation
+
+No formal paper citation is supplied with this snapshot. When using it, cite
+both this dataset repository and the pinned source repository, for example:
+
+```bibtex
+@misc{asi_bench_seed31415_2026,
+  title        = {ASI-Bench Generated Instances: Seed 31415 Snapshot},
+  author       = {{Apex Intelligence AI}},
+  year         = {2026},
+  howpublished = {Hugging Face dataset},
+  url          = {https://huggingface.co/datasets/Apexintelligence-AI/ASI-Bench-seed31415},
+  note         = {Generated from Agent-AI4Sci-Bench commit 6192234d4b9fdd8657a8054ac12b4cbaa827ea26}
+}
+```