CDR-Bench - Atomic Mappers: leaderboard

Metric: Recipe success RS@3 (%) on the atomic mapper (single text-rewriting operator) 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 2030. 11 models tracked.

Top models

#ModelScore
1GPT-5.4 (Non-reasoning)30.3
2Claude Opus 4.629.55
3Gemma 4 31B (IT) (Non-reasoning)29.55
4Kimi K2.6 (Non-reasoning)29.23
5Claude Opus 4.527.27
6GLM-5 (Non-reasoning)26.15
7Qwen 3.6 Max26.15
8DeepSeek V4 Flash (Non-reasoning)24.62
9Qwen 3.6 27B (Non-reasoning)24.24
10Qwen 3.6 35B A3B (Non-reasoning)18.18
11Llama 4 Scout16.92

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