mepi31415/WorldToken_Experiment_Records

WorldToken Experiment Records

Paper | Code | Experiment records | Checkpoints | Hugging Face paper page

Training logs, configurations, evaluation results and rollout videos for WorldToken: Time-First Sequence Modeling for Robotic Imitation Learning.

Paper section	Assets	Training runs
§4 — Multitask control and scaling	N1–N5 scaling grid, all five data settings, seeds 0/1; BC-Transformer baseline	50 + 1
§5 — Token interface	N2, K=4/K=50, all five data settings, seed 0	10
§6 — Recent history	N3, D300, C_train=1/2/5, seeds 0/1	6
§7 — Long history on RMBench	Nine-task initialization and modified-loss Blocks Ranking continuation	2

The package contains 69 training runs, 351 RoboCasa evaluation executions (403,650 episode executions), 500 formal RMBench episodes, nine exploratory continuation tests and 8,622 rollout MP4 videos. The supplementary galleries contain another 32 video files, for 8,654 MP4 files in the complete package.

Files within each run

• run_info.json: public run name, original training run name, parameters and final training step.
• config.json and configuration snapshots: recorded training and model settings.
• metrics.jsonl.gz: training and validation metrics; BC-Transformer uses log.txt.gz, and RMBench initialization also contains train.log.gz.
• Split files, dataset manifests and task conditions: recorded data selection for the corresponding runs.
• rollouts/<evaluation>/: episode results, summary statistics, evaluation settings and saved videos. Large text records use lossless gzip compression.
• RMBench diagnostics/: per-action-step evaluator records; behavior_analysis/: four analysis scripts. Continuation tests contain segment and event logs.

Run directory names encode the experiment and its parameters, for example scaling_n1_d300_seed0. Matching runs use the same directory names in the checkpoint package.

Aggregated evaluation settings

evaluation_protocols.json describes four shared protocols: RoboCasa WorldToken, RoboCasa BC-Transformer, RMBench formal evaluations, and RMBench exploratory continuation tests. evaluation_index.jsonl contains one sourced entry per evaluation directory: 351 RoboCasa evaluations, five formal RMBench evaluations, and nine exploratory RMBench trajectories, 365 entries in total.

These are release-time catalogues assembled from retained records and the cited paper/code snapshot. They are not newly recovered historical execution configs. The original files remain the underlying sources.

Each setting carries value, status, and sources:

Status	Meaning
recorded	Extracted from retained fields, including uniform or grouped episode metadata.
derived	Computed from records or directory naming; the derivation/source is identified.
protocol_declared	Specified by cited paper/code; not an original snapshot of every execution.
unknown	Not established from retained sources; null/empty is not proof of absence.

Shared declarations are stored once in the catalogue. Each index entry references its protocol_id; its settings override or supplement that profile. Source paths are relative to scope: records is this package, code is the WorldToken code repository, and paper is the paper source directory. source_files lists cited code/paper paths; source references use scope, path, and optional pointer, line, or symbol. For JSONL sources, pointer addresses a row field and selector states the relevant row scope.

The index preserves these distinctions:

• C_train is separate from C_test; holdout_eval_seed is separate from closed-loop rollout_seed or RMBench candidate seeds.
• WorldToken's RoboCasa observation/execution cadence is separate from the BC-Transformer baseline's native cadence and GMM sampler.
• RoboCasa runtime_exceptions preserve the affected summary's 27 rerun IDs and version/impact notes. The main runtime declaration is not a retained runtime snapshot for every episode.
• RMBench's selected task/seed subset is separate from its nine-task seed bank; the history cap is separate from realized trajectory history. Exploratory final status comes from final summary/status, not a startup running status.
• Archived shard/config paths in older metadata are provenance references; they are not assumed to be files included in this package.

Read resolved settings with Python's standard library, from this directory:

import json
from pathlib import Path

catalogue = json.loads(Path("evaluation_protocols.json").read_text(encoding="utf-8"))
wanted = "04_robocasa_scaling/scaling_n2_d300_seed0/rollouts/C1_repeat01"
with Path("evaluation_index.jsonl").open(encoding="utf-8") as handle:
    entry = next(item for item in map(json.loads, handle) if item["evaluation_id"] == wanted)
settings = dict(catalogue["protocols"][entry["protocol_id"]]["settings"])
settings.update(entry["settings"])
print({key: settings[key]["value"] for key in
       ("C_train", "C_test", "holdout_eval_seed", "rollout_seed",
        "observation_stride", "execute_horizon", "sampler")})

Download and use

Use the Hugging Face CLI to download the complete archive:

hf download mepi31415/WorldToken_Experiment_Records --repo-type dataset --local-dir ./WorldToken_Experiment_Records

Set RECORDS_ROOT to the absolute path of the downloaded WorldToken_Experiment_Records directory when running the paper summarizers. Matching released weights are in the companion checkpoint repository.

Download the repository files while preserving their directory structure. This package is an archive of experiment records with several JSON schemas, compressed logs and videos. Use the code repository's table reproduction instructions and section-specific summarizers to reproduce the reported tables. The original training demonstrations are obtained separately as described in the code repository's data preparation guide.

For the supplementary galleries, download the complete demo/ directory and open either gallery's index.html in a local browser; keep the adjacent video directories in place.

License

Original experiment records and documentation are released under CC BY 4.0. The RMBench behavior-analysis scripts have a separate MIT license. Third-party materials retain their respective licenses and attribution requirements.