Datasets:
language:
- en
task_categories:
- visual-question-answering
MemAct release
This prepared distribution contains 1,000,000 training QA turns with memory write/retrieval supervision, 14,001 RL tasks, and separate 20,000-question main and 2,000-question Timing evaluation banks. It includes self-rendered videos, generated Action/RL views, scene assets, and reconstruction scripts for real media obtained upstream. Model weights and original real recordings are excluded.
Follow RELEASE_GUIDE.txt and verify release_manifest.json before extracting.
All archive parts use prefixes under one data root. The extended frozen Action
bank has complete three-attempt branch coverage. Code and checks are supplied.
Read RELEASE_ALIGNMENT.txt for replacement-data statistics and method scope.
The native callback recomputes context across calls; it does not implement the
paper's persistent compact KV cache or reproduce reported latency. Different
real recordings can depict the same physical space; strict physical-space
separation is not claimed. Runtime checks do not reproduce historical scores.
This staging operation performs no upload. Rights/license decisions remain separate from technical integrity checks; no blanket license is granted here.