Add Control v2 incremental experiment results (evidence, visuals, 42 fixed samples)
a64600b verified |
Download method_v2/export_report.md from CharlesZhang-USTC/snf_experiments: direct link, hf CLI and curl.
- Browser
- Download file 1.97 kB
-
https://huggingface.co/CharlesZhang-USTC/snf_experiments/resolve/main/method_v2/export_report.md
- Command line
-
hf download hf://CharlesZhang-USTC/snf_experiments/method_v2/export_report.md
-
curl -L -o export_report.md https://huggingface.co/CharlesZhang-USTC/snf_experiments/resolve/main/method_v2/export_report.md
1.97 kB
| # Control v2 export | |
| Source: /jizhi/jizhi2/worker/trainer/snf/runtime/outputs/method-v2 | |
| Code actually run: /jizhi/jizhi2/worker/trainer/snf/snf-method-v2 (includes the numpy legacy-alias env fix in suite/momask_loading.py) | |
| ## Phases (gate: planned == completed, failed == 0) | |
| { | |
| "smoke": { | |
| "completed": 63, | |
| "curves": 21, | |
| "gallery": "/jizhi/jizhi2/worker/trainer/snf/runtime/outputs/method-v2/smoke/summary/index.html", | |
| "worker_failures": [], | |
| "planned": 63, | |
| "incomplete": 0 | |
| }, | |
| "dev": { | |
| "completed": 2772, | |
| "curves": 84, | |
| "gallery": "/jizhi/jizhi2/worker/trainer/snf/runtime/outputs/method-v2/dev/summary/index.html", | |
| "worker_failures": [], | |
| "planned": 2772, | |
| "incomplete": 0 | |
| }, | |
| "mechanism": { | |
| "completed": 216, | |
| "curves": 72, | |
| "gallery": "/jizhi/jizhi2/worker/trainer/snf/runtime/outputs/method-v2/mechanism/summary/index.html", | |
| "worker_failures": [], | |
| "planned": 216, | |
| "incomplete": 0 | |
| }, | |
| "test": { | |
| "completed": 15012, | |
| "curves": 2268, | |
| "gallery": "/jizhi/jizhi2/worker/trainer/snf/runtime/outputs/method-v2/test/summary/index.html", | |
| "worker_failures": [], | |
| "planned": 15012, | |
| "incomplete": 0 | |
| } | |
| } | |
| ## Contents | |
| - evidence/: numeric evidence for all phases, calibration, test-analysis, test-plan, plans, | |
| actual running code (suite/packaging/docs), locks (weights, env freeze, nvidia-smi, tasks sha) | |
| - paper-visuals/: fixed qualitative figures produced by the pipeline | |
| - samples/: 42 pre-fixed samples (first test task per model, seed=3, requested effect 0.5, | |
| raw+mapped x 7 methods). Selection ignores outcomes. | |
| - selected_jobs.csv, missing_files.csv, SHA256SUMS.txt | |
| ## Caveats | |
| - calibration_evaluation.json denominator is complete_locked_manifest: unreachable slots are kept | |
| in the denominator and were not inferred. | |
| - interpretation field: requested-effect evaluation; does not establish intrinsic generation | |
| capacity or causal mechanism. | |
| - MoMask did not improve after calibration; retained as a negative result. | |