|
Download requirement_matrix.md from LianeMarilin/ECG-EdgeBench: direct link, hf CLI and curl.
- Browser
- Download file 2.43 kB
-
https://huggingface.co/datasets/LianeMarilin/ECG-EdgeBench/resolve/main/requirement_matrix.md
- Command line
-
hf download hf://datasets/LianeMarilin/ECG-EdgeBench/requirement_matrix.md
-
curl -L -o requirement_matrix.md https://huggingface.co/datasets/LianeMarilin/ECG-EdgeBench/resolve/main/requirement_matrix.md
2.43 kB
Requirement freeze
| ID | Source | Exact requirement | Acceptance test | Status |
|---|---|---|---|---|
| R01 | EdgeBench integration guide | One task JSON under a benchmark containing BENCHMARK.yaml |
Load with the upstream load_task_spec implementation |
pass |
| R02 | EdgeBench task schema | Separate Work and Judge environments | Work payload contains no evaluator, private labels, or DS2 waveforms; Judge payload contains them | pass |
| R03 | EdgeBench task schema | Explicit task_id, name, base image, platform, cwd, submit paths, network policy, setup, prompt, evaluator, parser and selection |
Schema audit over generated task JSON | pass |
| R04 | EdgeBench parser contract | Judge prints a marked structured_json result |
Parser round-trip test | pass |
| R05 | User request | Real paper and real data | DOI, dataset version, license, record IDs and SHA-256 ledger are retained | pass |
| R06 | De Chazal et al. (2004) | Inter-patient DS1/DS2 evaluation on the 44 non-pacemaker MIT-BIH records | Exact frozen record lists; disjointness and 44-record union tests | pass |
| R07 | MIT-BIH v1.0.0 | Use two-channel, 360 Hz ECG and expert reference annotations | Download only from the versioned PhysioNet path; verify all expected .hea, .dat, .atr files |
pass |
| R08 | Scientific validity | Prevent beat-level/patient leakage | No record identifier occurs in both DS1 and DS2 | pass |
| R09 | High difficulty | Reward cross-patient minority-class performance, not majority accuracy | Composite includes macro F1, minority F1, minority recall geometry and record-level robustness | pass |
| R10 | Evaluation integrity | Submitted code cannot read hidden labels/evaluator | Private directory mode and unprivileged contestant subprocess test | pass |
| R11 | Iterative EdgeBench behavior | Continuous 0–100 score plus diagnostic metrics | selection=valid_then_score, score_direction=maximize, structured feedback |
pass |
| R12 | Reproducibility | Dataset construction is deterministic and auditable | Repeated build gives identical arrays and provenance hashes | pass |
| R13 | Package integrity | No secrets, absolute host paths, temporary files or .DS_Store |
Delivery validator and ZIP CRC/extraction check | pass |
Verdicts are reported separately as schema sample, coverage sample, and final delivery. A local task package is not an official EdgeBench task or leaderboard qualification.