---
license: other
task_categories:
- visual-question-answering
- reinforcement-learning
language:
- en
tags:
- gameworld
- multimodal-agents
- interaction-traces
- browser-games
pretty_name: GameWorld Interaction Traces
---

# GameWorld Interaction Traces

This dataset contains model interaction traces exported from local GameWorld
evaluation runs. Each row in `data/interactions.jsonl.gz` represents one model
decision step, including the prompt, model response, parsed action, runtime
state, task evaluation fields, and screenshot references available in the
original replay.

## Files

- `data/runs.jsonl.gz`: one row per rollout.
- `data/interactions.jsonl.gz`: one row per model interaction step.
- `images/screenshots-*.tar`: screenshot shards referenced by interaction rows.
- `metadata/export_manifest.json`: export counts and provenance.

## Export Summary

- Dataset repo target: `Fascetta/gameworld-interactions`
- Exported runs: 2067
- Exported interactions: 181447
- Exported interactions with screenshots: 152595
- Exported screenshots: 152595
- Exported at: 2026-09-21T15:43:49.382096+00:00

## Source

The traces were produced with the GameWorld benchmark runtime in this
repository. Raw local result directories are not required to load this dataset;
the source-relative paths are retained for provenance and auditability.

## Screenshot Layout

The interaction rows include `screenshot`, `screenshot_annotated`, and
`memory_screenshots` fields when those images were available locally. Each
image reference contains:

- `shard_path`: tar file inside the dataset repository.
- `member_path`: member path inside that tar file.
- `bytes`: original file size.
- `source_path`: original local path retained for auditability.

## Caveats

The dataset contains model prompts, screenshots, model outputs, parsed actions,
and runtime metadata from automated benchmark runs. It does not copy replay
videos, browser runtime folders, or local model weights.