| license: mit | |
| language: | |
| - en | |
| tags: | |
| - recap | |
| - robotics | |
| - video-prediction | |
| - tokenizer | |
| base_model: | |
| - thuml/rt1-compressive-tokenizer | |
| - thuml/rt1-world-model-multi-step-rlvr | |
| # rt1_tokenizer | |
| Dataset-specific tokenizer for ReCAP visual open-loop experiments. | |
| - Training: Published upstream starting checkpoint; training steps not independently verified. | |
| - Initialization: thuml/rt1-compressive-tokenizer (codec) and thuml/rt1-world-model-multi-step-rlvr (world model). | |
| - Format: diffusers; full strict-load weights, no optimizer/RNG state in inference release. | |
| - Pair only with `rt1_tokenizer`, `rt1_world`, and `rt1_action_ranges` from the same manifest. | |
| - Download: [matched model files](https://huggingface.co/easylearning/recap-robot-models/tree/61f3c350349f58ba8619c2454b5fac5802756282/weights/rt1_tokenizer). | |
| - License: MIT, retaining the starting model terms and attribution. This card does not grant rights over the training dataset. The Apache-2.0 code license is separate from checkpoint licensing. | |
| - Intended use: research video prediction with supplied actions. Not validated as a deployed robot controller. | |
| - Metrics: see artifacts/results; selected GIFs are not representative aggregate estimates. | |
| - Integrity: exact per-file SHA-256 and byte counts are in weights_manifest.json. | |
| ReCAP is training-free context scheduling and has no additional model parameters. | |