|
Download reports/experiments/availability.md from BurnyCoder/luna-folder-swarm-results: direct link, hf CLI and curl.
- Browser
- Download file 1.34 kB
-
https://huggingface.co/datasets/BurnyCoder/luna-folder-swarm-results/resolve/main/reports/experiments/availability.md
- Command line
-
hf download hf://datasets/BurnyCoder/luna-folder-swarm-results/reports/experiments/availability.md
-
curl -L -o availability.md https://huggingface.co/datasets/BurnyCoder/luna-folder-swarm-results/resolve/main/reports/experiments/availability.md
1.34 kB
Service availability and technical repeats
The primary analysis retains the recorded first attempts. The secondary calculations below use only questions with all five arms graded. The unaffected panel excludes a question from every arm when any generation failure logged an explicit provider overload. The repeat panel substitutes an eligible technical repeat only when it completed successfully. An unresolved repeat removes that question from every arm in that panel. This sensitivity analysis was specified after an upstream incident. Dated acquisition amendment
| Domain | Method | Primary complete pairs | Without flagged questions | With technical repeats |
|---|---|---|---|---|
| news | swarm4 | 74.6% (n=169) | 76.8% (n=164) | Postponed |
| news | swarm1 | 68.6% (n=169) | 70.7% (n=164) | Postponed |
| news | stock | 82.8% (n=169) | 85.4% (n=164) | Postponed |
| news | arctic | 78.1% (n=169) | 79.3% (n=164) | Postponed |
| news | oneshot | 71.0% (n=169) | 72.6% (n=164) | Postponed |
| code | swarm4 | 76.00 (n=81) | 77.88 (n=78) | Postponed |
| code | swarm1 | 74.53 (n=81) | 77.40 (n=78) | Postponed |
| code | stock | 79.27 (n=81) | 81.17 (n=78) | Postponed |
| code | arctic | 80.37 (n=81) | 82.35 (n=78) | Postponed |
| code | oneshot | 67.91 (n=81) | 68.59 (n=78) | Postponed |