Blind pairwise bucket-boundary calibration

Sets quality-bucket boundaries through blind pairwise comparisons of candidate thresholds.

source: 1
model: 1
lab adopt it: 1
strongest: used

How sources treat it

One count per evidence span, weakest treatment to strongest.

used 1

Documented in

Laguna-S-2.1

Evidence

1 span quoted from the sources, strongest treatment first.

For sampling, we define bucket boundaries by blind pairwise comparisons around candidate thresholds, so documents across a boundary are distinguishable by expected pre-training value.

useddata curationin Laguna XS.2Poolside

laguna m1 xs2 technical report·Poolside

Filed alongside

Other methods under data curation :: data mixture & curriculum.

AutoMixer Deterministic pre-splitting of ultra-long documents Vulnerability-discovery data inclusion Agent-centric multimodal data mixture Attribution, hedging, and refusal data subsets Data Scheduler Domain-specific multimodal datasets Dynamic sampling for mixed-task reinforcement learning Filtering, deduplication, and difficulty calibration Gaussian-based data mixture and curriculum construction Interleaved data Interleaved multimodal training KL-regularized data-mixture optimization Long-context data upsampling Pass-rate-based RL task sampling Prior-constrained Dirichlet mixture exploration ProtocolQA-aligned RL training datasets Sample Mixer Scaling-ladder-guided data and model planning Smooth weighted round-robin source scheduling Text-only pre-training corpus Three-stage data mixture strategy Three-stage pre-training curriculum Two-phase curriculum