Datasets:

mindrl-hub
/

mindrl-challenge-public

@@ -1,85 +1,185 @@
-# MindRL Challenge Public Dataset
-This repository contains the public training split for a human reward-learning
-benchmark. Participants repeatedly chose among four actions and received scalar
-reward feedback. Rewards were generated by latent, slowly changing payoff
-schedules, so good performance requires learning from recent outcomes while
-remaining responsive to change.
-The public release is designed for developing MindRL agents that predict human
-choices from trial history. It intentionally describes the task structure without
-identifying the original study or raw data source.
-## Task
-- Task family: 4-arm drifting bandit / restless reward-learning.
-- Action space: integers `0`, `1`, `2`, `3`.
-- Observation history: previous choices, received rewards, response times, and
-  anonymized episode metadata.
-- Prediction target: the human participant's next action.
-- Episode unit: each trajectory is one anonymized participant episode. The
-  anonymized `metadata.block_id` can be used to identify that two released
-  trajectory segments came from the same underlying episode, but it is not the
-  original block number.
-## Splits
-- `public_train.jsonl`: agent-facing public training trajectories. For each
-  released episode, the first portion of valid trials is released.
-- `public_train_reward_schedules.jsonl`: full option payoff schedules for the
-  released public episodes. This sidecar supports organizer-side generative
-  simulations and task reconstruction; it is not needed for standard
-  one-step-ahead choice prediction.
-Hidden evaluation splits are stored separately in the private MindRL Challenge
-dataset repository and are not part of this public release.
-## JSONL Format
-Each line is one trajectory:
-```json
-{
-  "context": {
-    "task_id": "public_4arm_restless_bandit",
-    "task_family": "restless_bandit",
-    "subject_id": "sub_000001",
-    "trajectory_id": "traj_000001",
-    "num_options": 4,
-    "available_actions": [0, 1, 2, 3],
-    "option_features": null,
-    "task_description": "4-arm drifting bandit task",
-    "metadata": {"block_id": "block_000001"}
-  },
-  "trials": [
-    {"trial_index": 0, "action": 1, "reward": 72.0, "info": {"rt": 531.0}}
-  ]
-}
-```
-Field notes:
-- `subject_id`, `trajectory_id`, and `metadata.block_id` are anonymized labels.
-- `trial_index` is the within-trajectory index after preprocessing.
-- `action` is the observed human choice.
-- `reward` is the feedback received after that choice.
-- `info.rt` is response time when available.
-## Safety / leakage
-- Original participant identifiers, original block numbers, original trial ids,
-  and raw study/source identifiers are not included.
-- Demographic variables are not included.
-- Counterfactual option payoffs are not included in the agent-facing trajectory
-  history. For public episodes, they are kept only in the separate
-  reward-schedule sidecar for generative simulation.
-- The dataset card intentionally omits the original article and raw-data source
-  so that challenge participants cannot trivially recover hidden evaluation
-  splits from external copies of the raw dataset.
-## Intended Use
-Use this public split to develop, debug, and calibrate MindRL submissions. The
-official challenge evaluation uses additional private splits to test
-generalization across held-out trials, participants, episodes, and a separate
-held-out task.

+---
+pretty_name: MindRL Challenge Public Dataset
+task_categories:
+- reinforcement-learning
+tags:
+- behavioral-modeling
+- reward-learning
+- bandit-task
+- human-choice-prediction
+language:
+- en
+---
+# MindRL Challenge Public Dataset
+This repository contains the public training split for the MindRL Challenge, a human reward-learning modeling benchmark. The release is intentionally source-safe: it describes the task, data format, and intended use without identifying the original study, raw data source, article, repository, lab, institution, or participant pool.
+## Intended use
+Use this public split to develop, debug, and calibrate MindRL agents that predict human choices from trial history. Official challenge scoring uses additional private evaluation splits that are not part of this public release.
+The modeling target is one-step-ahead behavioral prediction:
+```text
+P(action_t | context, history_1:t-1)
+```
+At prediction time, models should use only the trajectory context and previous trials. They should not use the current trial's action or reward, later trials from the same trajectory, hidden splits, or any external reconstruction of the raw data source.
+## Task
+- Task family: 4-arm drifting bandit / restless reward-learning.
+- Action space: integers `0`, `1`, `2`, `3`.
+- Observation history: previous choices, received rewards, response times when available, and anonymized episode metadata.
+- Prediction target: the human participant's next action.
+- Episode unit: each JSONL row is one anonymized trajectory.
+- The reward associated with each action changes gradually over the episode, so useful models need to learn from recent outcomes while remaining responsive to change.
+## Files
+| File | Purpose |
+| --- | --- |
+| `public_train.jsonl` | Agent-facing public training trajectories. |
+| `public_train_reward_schedules.jsonl` | Full option payoff schedules for released public episodes; useful for public-task analysis or generative simulation, but not part of the standard one-step-ahead observation history. |
+| `schema.json` | JSON Schema for public trajectories. |
+| `task_description.md` | Compact task description. |
+| `README.md` / `dataset_card.md` | This dataset description. |
+If a model uses the reward-schedule sidecar, that use should be disclosed in the method description or interpretation card. The official prediction API still evaluates one-step-ahead predictions from context and past observed trials.
+## Public split statistics
+These statistics describe `public_train.jsonl` only.
+| Quantity | Value |
+| --- | ---: |
+| Trajectories | 2,678 |
+| Anonymized subjects | 689 |
+| Anonymized blocks / episodes | 2,678 |
+| Total trials | 320,080 |
+| Trajectory length, min / median / max | 108 / 120 / 120 |
+| Reward range | 1.0 to 100.0 |
+| Mean reward | 64.889 |
+| Trials with response time | 308,803 |
+Observed action counts:
+| Action | Count |
+| --- | ---: |
+| `0` | 81,656 |
+| `1` | 81,698 |
+| `2` | 77,045 |
+| `3` | 79,681 |
+These counts should not be interpreted as the full challenge distribution.
+## JSONL format
+Each line in `public_train.jsonl` is one trajectory:
+```json
+{
+  "context": {
+    "task_id": "public_4arm_restless_bandit",
+    "task_family": "restless_bandit",
+    "subject_id": "sub_000001",
+    "trajectory_id": "traj_000001",
+    "num_options": 4,
+    "available_actions": [0, 1, 2, 3],
+    "option_features": null,
+    "task_description": "4-arm drifting bandit task",
+    "metadata": {"block_id": "block_000001"}
+  },
+  "trials": [
+    {"trial_index": 0, "action": 1, "reward": 72.0, "info": {"rt": 531.0}}
+  ]
+}
+```
+### Context fields
+| Field | Meaning |
+| --- | --- |
+| `task_id` | Public task identifier: `public_4arm_restless_bandit`. |
+| `task_family` | Coarse task family: `restless_bandit`. |
+| `subject_id` | Anonymized subject label such as `sub_000001`; not an original participant id. |
+| `trajectory_id` | Anonymized trajectory label such as `traj_000001`. |
+| `num_options` | Number of available actions. |
+| `available_actions` | Valid actions, encoded as integers `[0, 1, 2, 3]`. |
+| `option_features` | Optional action-feature field; `null` in this release. |
+| `task_description` | Short task description string. |
+| `metadata.block_id` | Anonymized episode/block label; not the original block number. |
+### Trial fields
+| Field | Meaning |
+| --- | --- |
+| `trial_index` | Within-trajectory trial index after preprocessing. |
+| `action` | Observed human action, encoded as `0`, `1`, `2`, or `3`. |
+| `reward` | Scalar reward feedback received after the action. |
+| `info.rt` | Response time when available. Some trials may have missing response time. |
+## Loading examples
+Download the public trajectory file:
+```bash
+huggingface-cli download mindrl-hub/mindrl-challenge-public public_train.jsonl --local-dir ./hf_cache/public
+```
+Load with Python JSONL:
+```python
+import json
+from pathlib import Path
+path = Path("hf_cache/public/public_train.jsonl")
+with path.open("r", encoding="utf-8") as f:
+    first = json.loads(next(f))
+print(first["context"]["available_actions"], len(first["trials"]))
+```
+Load with `datasets`:
+```python
+from datasets import load_dataset
+ds = load_dataset(
+    "json",
+    data_files="https://huggingface.co/datasets/mindrl-hub/mindrl-challenge-public/resolve/main/public_train.jsonl",
+    split="train",
+)
+print(ds[0])
+```
+## Evaluation interface
+The official evaluator calls an `Agent` sequentially:
+1. `reset(context)` once at the start of a trajectory.
+2. For each trial, call `predict(history)` using only previous trials.
+3. Score the predicted action distribution against the observed human action.
+4. Call `update(action, reward, info)` after the true trial outcome is revealed.
+5. Append the trial to history and continue.
+A valid prediction returns probabilities over the available actions, for example:
+```python
+{"action_probs": {0: 0.25, 1: 0.25, 2: 0.25, 3: 0.25}}
+```
+## Safety and leakage policy
+- Original participant identifiers, original block numbers, original trial ids, raw filenames, and raw study/source identifiers are not included.
+- Demographic variables are not included.
+- `subject_id`, `trajectory_id`, and `metadata.block_id` are anonymized labels.
+- Counterfactual option payoffs are not included in the agent-facing trajectory history. For public episodes, they are kept only in the separate reward-schedule sidecar.
+- The dataset card intentionally omits the original article and raw-data source so that challenge participants cannot trivially recover hidden evaluation splits from external copies of the raw dataset.
+Do not use this release to identify the original study, raw dataset, collection site, institution, lab, article, or participant pool. Do not add source-identifying details to public repositories, participant documentation, or submissions.
+## Relationship to hidden evaluation
+The public split supports development. Official challenge evaluation uses hidden splits prepared by the organizers to test generalization across unreleased trials, unreleased participants, unreleased episodes, and a separate held-out task. Hidden evaluation files and source mappings are not part of this public release.