implementation detail · filed under data curation
KL-regularized data-mixture optimization
Regularizes mixture optimization toward a baseline prior to avoid unrealistic concentration in a few data sources.
- 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
Evidence
1 span quoted from the sources, strongest treatment first.
To avoid unrealistic shifts toward a small number of dominant sources, we regularize the optimization toward the baseline prior
useddata curationin Laguna XS.2Poolside
Filed alongside
Other methods under data curation :: data mixture & curriculum.
AutoMixerDeterministic pre-splitting of ultra-long documentsVulnerability-discovery data inclusionAgent-centric multimodal data mixtureAttribution, hedging, and refusal data subsetsBlind pairwise bucket-boundary calibrationData SchedulerDomain-specific multimodal datasetsDynamic sampling for mixed-task reinforcement learningFiltering, deduplication, and difficulty calibrationGaussian-based data mixture and curriculum constructionInterleaved dataInterleaved multimodal trainingLong-context data upsamplingPass-rate-based RL task samplingPrior-constrained Dirichlet mixture explorationProtocolQA-aligned RL training datasetsSample MixerScaling-ladder-guided data and model planningSmooth weighted round-robin source schedulingText-only pre-training corpusThree-stage data mixture strategyThree-stage pre-training curriculumTwo-phase curriculum