yyshi0619/ECOT-Alignment-900-Episodes

ECOT Alignment 900 Episodes

This dataset contains 900 rollout episodes generated by the original MiniVLA policy across all 90 LIBERO-90 tasks (10 distinct initial configurations per task). It was collected for supervised policy/reasoning alignment experiments. Successful and failed episodes are both included.

Splits and counts

Split	Episodes	Policy queries
Training	720	13,093
Validation	180	3,229
Total	900	16,322

The split is task-stratified: each task contributes eight training episodes and two validation episodes. Exact episode IDs are recorded in assembled/split.json. Each row in assembled/records.jsonl also has a split field.

Repository layout

• assembled/records.jsonl: the complete query-level fine-tuning dataset.
• assembled/episodes.json: episode-level metadata and outcomes.
• assembled/split.json: exact training and validation episode IDs and counts.
• assembled/dataset_complete.json: integrity totals.
• assembled/dataset_statistics.json: alignment statistics.
• assembled/plan.json: frozen collection and experiment configuration.
• raw/large-data-collection-00 through 08: all raw rollout outputs, frames, videos, action/reasoning traces, summaries, and per-rank shards for 100 episodes each.
• provenance/: collection commands and the deterministic assembly script.

Loading the assembled records

from datasets import load_dataset

ds = load_dataset(
    "json",
    data_files="hf://datasets/yyshi0619/ECOT-Alignment-900-Episodes/assembled/records.jsonl",
    split="train",
)
train = ds.filter(lambda row: row["split"] == "training")
validation = ds.filter(lambda row: row["split"] == "validation")

The before_frame values preserve the absolute paths on the collection workstation. On another machine, replace the prefix /home/exx/Projects/ECOT-Alignment/experiments/robot/libero/results/ with the local path to this repository's raw/ directory. The remaining suffix starts with large-data-collection-XX/... and is preserved exactly.

Provenance and intended use

The supervision consists of original-MiniVLA-generated reasoning and action tokens paired with the policy's executed LIBERO rollouts. This is suitable for research on VLA fine-tuning, reasoning/action alignment, and failure analysis. It should not be interpreted as human-authored reasoning or real-robot data.

Users are responsible for following the licenses and terms of the upstream MiniVLA/OpenVLA and LIBERO assets used to produce these rollouts.