rogercc/agentick-oracle-trajectories-250k

Agentick Oracle Trajectories — 250k

Synthetic, per-step demonstrations generated by Agentick's programmatic oracle policies for research on sequential decision-making agents. MIT licensed. Maintained by Roger Creus Castanyer; associated paper authors are Roger Creus Castanyer, Pablo Samuel Castro and Glen Berseth.

• Code and oracle policies
• Documentation
• Croissant metadata with Responsible AI fields
• Measured dataset inventory and file checksums
• License

Contents and sizes

One row is one environment step, not a complete episode. The 250k label approximates the number of training rows. Counts below were measured from every released Parquet shard on October 16, 2026. Unique trajectories are counted by (task, difficulty, episode_id) within a split.

Split	Rows / steps	Trajectories	Task IDs	Difficulties
train	251,076	3,800	38	4
test	252,123	3,800	38	4

Each task/difficulty combination has [25] training trajectories and [25] test trajectories. All four difficulties (easy, medium, hard, expert) are represented. The files contain 38 task IDs, including CausalChain-v0, which is not in the paper's 37-task list. The historical data are preserved; filter this additional task when selecting the paper's task set. Record that filtering in experiment reports.

Loading and splits

from datasets import load_dataset

ds = load_dataset("rogercc/agentick-oracle-trajectories-250k", revision="58c01a6f17afbfbfc15165c6ced2d0e71a5f5135")
train = ds["train"]
test = ds["test"]
# Optional: select the 37-task set listed in the paper.
train_paper_tasks = train.filter(lambda row: row["task"] != "CausalChain-v0")

Use train for training and test only for held-out analysis. The public collector uses train and eval seed namespaces, respectively. Do not train on test trajectories or use them for hyperparameter selection before claiming held-out benchmark performance. Dataset variants can share trajectories; do not concatenate them as independent data. The 250k and 500k test splits have identical row counts but different file hashes; this metadata update does not establish whether their individual trajectories overlap.

The revision above pins the original data snapshot. The current main branch adds documentation and metadata; Parquet bytes are unchanged. Croissant record sets correspond to individual physical shards and are named by split and shard, so loaders can iterate every shard without an implicit repository glob.

Schema

Field	Stored type	Meaning
task	Text	Environment task identifier, including version suffix.
episode_id	Int32	Episode identifier. Group with task, difficulty and split; sort by step to reconstruct a trajectory.
difficulty	Text	Procedural difficulty: easy, medium, hard or expert.
step	Int32	Zero-based step index within an episode.
ascii_render	Text	ASCII observation before the action; may contain ANSI colour escape sequences.
language_render	Text	English template-generated observation before the action.
action_name	Text	Oracle action name in the task-specific action space.
action_int	Int32	Oracle action index in the task-specific action space; interpret together with task.
reward	Float32	Environment reward returned after the action.
done	Boolean	True when the environment terminated or truncated after the action; these two cases are combined.

Action indices are task-specific; do not assume a single index-to-action mapping for every task. done combines termination and truncation. Observations precede the action, whereas reward and done describe its outcome. The files contain text observations, not pixels, chain-of-thought rationales or human annotations.

Generation and provenance

Programmatic Agentick oracles interact with procedurally generated environments. Each row records pre-action ASCII and language observations, an oracle action and the resulting reward and termination-or-truncation flag. The public collector is documented at https://github.com/roger-creus/agentick/blob/279fe5f34a35196ba3904550f911d5c8ade3c7c7/examples/data_and_finetuning/collect_oracle_trajectories.py; the documented seed recipe is at https://github.com/roger-creus/agentick/blob/279fe5f34a35196ba3904550f911d5c8ade3c7c7/agentick/leaderboard/seeds.py. This pinned source snapshot describes the current public pipeline; the exact commit used to produce the historical files is not recorded, and byte-identical regeneration is not claimed.

The inspected collector appends each interaction step, casts values to string/int32/float32/bool, and serializes train/test splits to Hugging Face Parquet. It does not filter episodes by success; environments that cannot initialize or lack an oracle may be skipped. No additional historical preprocessing is recorded. Data files are unchanged in this metadata release; each file is pinned by original dataset revision and SHA-256.

The documented seed function hashes "{task_name}::{difficulty}::{split}" with SHA-256, interprets the first 16 hexadecimal characters as an integer, seeds NumPy default_rng, and draws integer seeds in [0, 2**31). The collector takes the first requested seeds for each task/difficulty pair. The raw per-row environment seed is not stored in this schema; episode IDs alone should not be interpreted as random seeds. Original generation timestamps and the precise historical code revision are unavailable in the released metadata. This card describes that limitation rather than inventing them.

Original data revision: 58c01a6f17afbfbfc15165c6ced2d0e71a5f5135. Documentation source snapshot: 279fe5f34a35196ba3904550f911d5c8ade3c7c7. File hashes and task inventories are in dataset_stats.json and croissant.json.

Responsible AI

Intended uses and validity

Intended for supervised fine-tuning, behaviour cloning, trajectory analysis and within-Agentick diagnostics. The paper reports supervised adaptation results; this does not establish validity for real-world safety, fairness, human intelligence or general intelligence measurement. Use only the train split for training. The public collector uses the benchmark eval seed namespace for the test split: do not train on test trajectories or use them for tuning before reporting held-out benchmark results.

Limitations and uses not established

Synthetic, discrete, two-dimensional, turn-based gridworld trajectories. Not representative of real-world continuous control, human behaviour, or unconstrained agents. Oracle policies are programmatic reference policies and are not guaranteed globally optimal. English and ASCII observations dominate; no demographic or real-world safety validity is established. Released files include CausalChain-v0 in addition to the 37 tasks listed in the paper; filter this task for the paper task set. Training-size labels count approximate training rows (steps), not episodes. Exact historical generator commit and collection dates are not recorded in the released files.

Biases

Task selection and hand-built oracle policies determine the behaviour distribution. Longer episodes contribute more rows, so uniform row sampling overweights long trajectories. Capability groupings are diagnostic categories, not a validated cognitive taxonomy. The synthetic worlds and English templates do not represent human demographic populations. Dataset variants may overlap; they must not be treated as independent samples or concatenated without overlap checks.

Personal or sensitive information

The documented generation process uses synthetic environments, programmatic oracle actions and template-generated observations, with no human participants or scraped personal records. No personal or sensitive information is intended to be included; episode IDs are synthetic trajectory identifiers, not person identifiers.

Annotation

Oracle policies supply action labels; environment execution supplies rewards and end-of-episode flags. There are no human annotators, so demographic participation and inter-annotator agreement are not applicable. Oracle trajectories may be suboptimal or unsuccessful; done does not certify success.

Potential social impact

May lower the cost of controlled agent research and reproducible comparisons. Improved agent capabilities can also be misused, and strong gridworld scores can be overinterpreted as evidence of real-world competence or safety. Mitigations are explicit task/split documentation, provenance, limitations, open inspection and keeping evaluation trajectories out of training. No safety certification is implied.

License and attribution

The dataset and accompanying metadata/documentation are released under the MIT License, matching the project code license, as confirmed by the dataset maintainer. Preserve the copyright and license notice when redistributing covered material. Third-party software retains its own licenses.

Associated paper: Agentick: A Unified Benchmark for General Sequential Decision-Making Agents (Roger Creus Castanyer, Pablo Samuel Castro and Glen Berseth, 2026). Please cite the final paper when available and include this dataset's URL and revision in experiments.

Related releases and contact

• 120k
• 250k
• 500k

Report issues via the Agentick repository or this dataset's Community tab. Metadata update: October 16, 2026. No regeneration or modification of the data files was performed for this update.