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

#ModelScore
1GPT-5.4 (Non-reasoning)88.1
2Qwen 3.6 27B (Non-reasoning)80.95
3Gemma 4 31B (IT) (Non-reasoning)80.95
4Qwen 3.6 35B A3B (Non-reasoning)78.05
5Kimi K2.6 (Non-reasoning)76.19
6DeepSeek V4 Flash (Non-reasoning)76.19
7Qwen 3.6 Max76.19
8Claude Opus 4.673.81
9GLM-5 (Non-reasoning)73.81
10Claude Opus 4.571.43
11Llama 4 Scout62.5

Interactive version: theaggregate.ai/benchmark?slug=cdr-bench-atomic-filters · How It Works · Data refreshed daily, snapshot 2026-09-29.