CDR-Bench - Atomic Filters: leaderboard
CDR-Bench (2026) tests whether LLMs faithfully execute data-refinement recipes written in natural language over web, LaTeX, RAG-preparation and privacy text, using 29 Data-Juicer operators with deterministic reference outputs. This board is the percent of atomic filter tasks (one KEEP or DROP statistical filter) that the model gets exactly right within three attempts, with each model in non-thinking mode. Higher is better.
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: Not forecast. 11 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | GPT-5.4 (Non-reasoning) | 88.1 | #123 (GPT-5.4) |
| 2 | Qwen 3.6 27B (Non-reasoning) | 80.95 | #341 (Qwen 3.6 27B) |
| 3 | Gemma 4 31B (IT) (Non-reasoning) | 80.95 | #342 (Gemma 4 31B (IT)) |
| 4 | Qwen 3.6 35B A3B (Non-reasoning) | 78.05 | #462 (Qwen 3.6 35B A3B) |
| 5 | DeepSeek V4 Flash (Non-reasoning) | 76.19 | #258 (DeepSeek V4 Flash) |
| 6 | Kimi K2.6 (Non-reasoning) | 76.19 | #178 (Kimi K2.6) |
| 7 | Qwen 3.6 Max | 76.19 | #180 |
| 8 | Claude Opus 4.6 | 73.81 | #102 |
| 9 | GLM-5 (Non-reasoning) | 73.81 | #260 (GLM-5) |
| 10 | Claude Opus 4.5 | 71.43 | #156 |
| 11 | Llama 4 Scout | 62.5 | #1776 |
No result here: #1 Claude Fable 5.2, #2 Gemini 4 Carbon, #5 Claude Opus 5.5.
Interactive version: theaggregate.ai/benchmark?slug=cdr-bench-atomic-filters · How It Works · Data refreshed daily, snapshot 2026-11-13.