implementation detail · filed under data curation

Pass@100 filtering

Synthetic RL tasks or environments are retained only when they have non-zero pass@100.

sources: 2
model: 1
lab adopt it: 1
strongest: used

How sources treat it

One count per evidence span, weakest treatment to strongest.

used 2

Documented in

DeepSeek-V3.2

Evidence

2 spans quoted from the sources, strongest treatment first.

We then perform RL on this dataset using DeepSeek-V3.2 and retain only instances with non-zero pass@100, resulting in 1,827 environments and their corresponding tasks (4,417 in total).

useddata curationin DeepSeek-V3.2DeepSeek

paper·DeepSeek

We then perform RL on this dataset using DeepSeek-V3.2 and retain only instances with non-zero pass@100

useddata curationin DeepSeek-V3.2DeepSeek

DeepSeek-V3.2: Pushing the Frontier of Open Large Language Models·DeepSeek·2025-12-02

Filed alongside

Other methods under data curation :: data filtering.

CSAM filtering Sensitive data filtering CBRN pre-training data filtering Filtering batched auto-generated and templated content Heuristic filtering Keyword- and regex-based filtering Model-based refusal and inability filtering Pathological repetition filtering SmolVLM-based image-text quality scoring Unified data filtering pipeline Automated verification Coding-agent session filtering and trajectory deduplication Conservative model-based noise filtering Content quality and safety filtering Continuous contribution-score ranking Data cleaning Data filtering pipeline Dense annotation of ambiguous low-quality data Domain filtering with heuristics, classifier scoring, and deduplication Evidence-based data cleaning and training constraints Filtering low-information model-generated content Heuristic and statistical filtering with deduplication and quality models Image-text relevance filtering Long-context data cleaning pipeline