FLOWSTEP Judge One-Shot - Krippendorff Alpha: leaderboard
FLOWSTEP v1 evaluates next-step prediction for GitHub Actions. It uses a repository-level 8:1:1 split and 13630 full test steps. Abs and Raw are distinct abstraction conditions. Generation uses temperature 0 and one candidate; calibrated script-model selection uses 1000 random test Script instances. The full pipeline combines RoBERTa type/action heads with Sonnet 4.5 one-shot prompting or a trained T5-small; the weights are trained on distinct task sets. Judge validation uses 350 incorrect scripts, 2 human annotators, and 3 ordinal functional classes. Signed agreement and correlation measure judgment quality, not generation accuracy. BLEU/ROUGE metric controls are not model rows. Judge endpoint checkpoints are unreported beyond the source labels. Metric: Krippendorff Alpha.
Metric: Krippendorff Alpha. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 5 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | FLOWSTEP judge GPT-4.1 One-Shot | 0.69 | |
| 2 | FLOWSTEP judge Claude-Sonnet-4.5 One-Shot | 0.66 | |
| 3 | FLOWSTEP judge DeepSeek-V3.2-Exp One-Shot | 0.66 | |
| 4 | FLOWSTEP judge Claude-Haiku-4.5 One-Shot | 0.61 | |
| 5 | FLOWSTEP judge GPT-3.5-turbo One-Shot | 0.28 |
Interactive version: theaggregate.ai/benchmark?slug=flowstep-judge-one-shot-krippendorff-alpha · How It Works · Data refreshed daily, snapshot 2026-10-28.