Model: mouse-dit-fm-20261024-v1
Architecture: conditional flow-matching Transformer, 6 blocks, width 192, 6 heads, 4,174,850 parameters.
Sources: user_mousedata, bogazici, sapimouse; source/task labels and normalization retained.
Training: independent implementation, trained from scratch; no previous user checkpoint reused.
Inference: ONNX Runtime CPU, Heun integration; predicted total duration and median cadence.
Limitations: native inter-event jitter and target width are not learned; source pixels are not calibrated across devices.
User drag holdout does not establish unseen-person or unseen-session generalization.
User drag energy score approximately tied with empirical sampling baseline; small backtracking remains elevated.
Source datasets are not included in this runtime package. No uniform dataset redistribution license is asserted.
Architecture references: https://arxiv.org/abs/2210.02747 ; https://arxiv.org/abs/2212.09748
