The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
auto1097 training data
This repository contains the 176 fixed training shards used by the auto1097
task: 180,000 video segments, 70 frames each, with actions. The files are the
original prepared NPZ bytes used by the recorded experiments; they have not
been regenerated or re-encoded. data_manifest.json records the version,
total size, and SHA-256 of every shard.
Each shard contains frames (uint8, N x 70 x 32 x 32), actions (int8,
N x 70 x 2), and original relative names. The data root contains train/.
Download a pinned repository commit before running the task. Mount only its
train/ subdirectory read-only at $AUTORESEARCH_DATA_ROOT/train in both
execution environments. Repository documentation and metadata are for the
platform preparation step; they are not mounted into the Agent. Evaluation
runs offline.
Source and preparation
Source: https://huggingface.co/datasets/flowm123/mnist-world, dynamic_po, dynamic_training. The source videos were decoded with PyAV, converted to grayscale using 0.2989 R + 0.5870 G + 0.1140 B, represented as uint8 at 32 x 32, and paired with the original int8 actions. Episode and shard order are preserved by the prepared data used in the task.
The upstream dataset card did not declare a dataset license when checked on 2026-10-08. This card does not assign a new license to the upstream data or infer a dataset license from the associated code license.
Only the training split is included. Dev, Hidden, task solutions, evaluation code, expert evidence, model weights, and credentials are outside this repository.
- Downloads last month
- 234