yor-icl-expanded-fast-tokenizer

A FAST (Frequency-space Action Sequence Tokenization) action tokenizer, fit on icl-dataset's expanded 31-task, 1,784-episode bimanual pick-and-place subset. Used by every pi0-fast/FAST-auxiliary-loss config in this repo trained on the expanded task set:

• yor_icl_pi0_fast_expanded / yor_icl_pi0_fast_expanded_continued_50k
• yor_icl_fast_victr_vision_expanded / _serve / _interp_expanded
• yor_icl_ki_subtask (auxiliary FAST loss branch, expanded variant)

Fit stats (metadata.json):

• 31 tasks, 1,784 episodes, 96,010 chunks
• action_horizon=30, action_dim=20
• vocab_size=1024, scale=10.0, sample_fraction=0.1
• normalization_mode=QUANTILES
• Compression ratio ≈ 20.8x (mean token length 28.8, p99 61.0)

Load with transformers.AutoProcessor (processor_config.json sets auto_map to processing_action_tokenizer.UniversalActionProcessor) or point openpi's Pi0FASTConfig/Pi0KiConfig fast_tokenizer_path at a local clone of this repo.