single mention · data curation
Joint refinement on embodied and general multimodal data
Joint refinement is a second training stage that continues fine-tuning an embodied-specialized model on a mixture of embodied data and general multimodal data.1 It balances the two data sources to trade off embodied-reasoning gains against retention of general capabilities.2
How this definition was written
Summarised by glm-5.3-flash from 1 source; every sentence is checked against the quote it cites.
We continue training the Stage 1 checkpoint for K additional steps on a mixture that interleaves our embodied corpus with Molmo2’s original multimodal mid-training data
— MolmoAct2 Action Reasoning Models for Real-World Deploymentyields the best Pareto trade-off between the embodied-reasoning benchmarks in Table 3 and Molmo2’s general benchmarks
— MolmoAct2 Action Reasoning Models for Real-World Deployment
Also called Joint refinement.
Filed under data curation › data mixture & curriculum