simate-beta initialization
Native FP32 initialization for the simate-beta visual-memory policy. This is an initialization checkpoint for training, not a final evaluation checkpoint.
It combines a pretrained EMA memory policy with a pretrained frame-wise history encoder. There are 701 tensors (667 policy tensors and 34 history-encoder tensors), containing 3,374,422,288 parameters. All source tensor values are preserved exactly. No optimizer state, training data, or AE decoder is included.
Files
- model.safetensors: complete native model initialization.
- tokenizer.model: matching SentencePiece tokenizer.
- state_action_statistics.json: matching frozen normalization statistics.
- SHA256SUMS: checksums for all assets.
Set SIMATE_INIT_CHECKPOINT, SIMATE_TOKENIZER, and SIMATE_STATISTICS to these files when using the SIPAI simate-beta training configuration. This is the native SIPAI parameter layout, not an AutoModel checkpoint.
Dataset manifests, phase annotations, and target overrides must be supplied separately. The normalization statistics retain their original phase fingerprint; configure the matching supervision assets and validated phase rebinding when required.
See LICENSE, LICENSE_OPENPI, and NOTICE for upstream terms and attribution.