CDR-Bench - Atomic Filters: leaderboard
Metric: Recipe success RS@3 (%) on the atomic filter (single KEEP/DROP statistical filter) tasks of CDR-Bench (3,462 data-refinement tasks over web, LaTeX, RAG-preparation and privacy text with 29 Data-Juicer operators and deterministic reference outputs): a task counts when the model's refined text or KEEP/DROP judgment exactly matches the reference within 3 attempts; non-thinking mode, baseline prompt; higher is better. Source: arxiv.org. Saturation forecast: Around December 2026. 11 models tracked.
Top models
| # | Model | Score |
|---|---|---|
| 1 | GPT-5.4 (Non-reasoning) | 88.1 |
| 2 | Qwen 3.6 27B (Non-reasoning) | 80.95 |
| 3 | Gemma 4 31B (IT) (Non-reasoning) | 80.95 |
| 4 | Qwen 3.6 35B A3B (Non-reasoning) | 78.05 |
| 5 | Kimi K2.6 (Non-reasoning) | 76.19 |
| 6 | DeepSeek V4 Flash (Non-reasoning) | 76.19 |
| 7 | Qwen 3.6 Max | 76.19 |
| 8 | Claude Opus 4.6 | 73.81 |
| 9 | GLM-5 (Non-reasoning) | 73.81 |
| 10 | Claude Opus 4.5 | 71.43 |
| 11 | Llama 4 Scout | 62.5 |
Interactive version: theaggregate.ai/benchmark?slug=cdr-bench-atomic-filters · How It Works · Data refreshed daily, snapshot 2026-09-29.