# Context Window: Flappy Bird

80 archived training runs. Seeds 10–14. Native Sample Factory APPO policies.

## Settings

Context windows: 1–8. Environment FPS: 30. Fixed delay: 2 environment frames. Uniform delay: 1–3 environment frames. Appendix evaluation horizon: 36,000 steps.

The archive retains each evaluation's original horizon, including earlier 3,600-step records. Each original evaluation has 100 episodes.

Exact observation cadence, frame stack, training budget and delay parameters belong to each run's unchanged `config.json` and [manifest](manifest.json). Fixed and uniform distributions are experiment settings, not measured hardware profiles.

## Results and assets

[Paper measurements](https://huggingface.co/datasets/MLL-Lab/LAGEN-datasets/blob/3a59b67c3f85954742905d2c57b0d40e0050f979/experiments/context-window/paper/data.csv) and [fixed W&B sources](https://huggingface.co/datasets/MLL-Lab/LAGEN-datasets/blob/3a59b67c3f85954742905d2c57b0d40e0050f979/experiments/context-window/paper/wandb_sources.csv) retain the published values. Each plotted point aggregates five training seeds.

The manifest pins model, configuration and original evaluation assets to fixed main-repository revisions. Training logs are stored with the evaluation run under `training/`. Download the model bundle with its original configuration and load it with the task's native Sample Factory integration.

## Limits

Retained run configurations and evaluation records preserve their original horizons. A matching training condition alone does not prove that two evaluation records are identical.