README.md

metadata

license: mit
language:
  - en
task_categories:
  - reinforcement-learning
  - text-classification
tags:
  - agent-routing
  - appworld
  - scienceworld
  - multi-model-routing

RouteStep Trajectory Dataset

This directory contains normalized AppWorld and ScienceWorld trajectories for step-level model routing. Every JSON file is self-contained and uses USD cache-aware token-price estimates from pricing.json.

Layout

dataset/
├── pricing.json
├── appworld/
│   ├── manifest.json
│   ├── task_ids.txt
│   ├── single_model_trajectories/<task_id>/<model_slug>.json
│   └── mixed_trajectories/<task_id>.json
└── scienceworld/
    ├── manifest.json
    ├── task_ids.txt
    ├── single_model_trajectories/<task_id>/<model_slug>.json
    └── mixed_trajectories/<task_id>.json

Cost convention

prompt_tokens includes cached tokens. For every model call:

uncached_input_tokens = prompt_tokens - cache_read_tokens - cache_write_tokens
estimated_cost_usd = (
    uncached_input_tokens * input_price
    + cache_read_tokens * cache_read_price
    + cache_write_tokens * cache_write_price
    + completion_tokens * output_price
) / 1_000_000

Provider-reported spend is not used for cross-model comparisons. Prices and all costs are denominated in USD.

In AppWorld single-model files, state.history_step_indices references earlier steps in the same trajectory instead of duplicating their complete action and observation payloads. This preserves the full ordered history while keeping the dataset compact.

Coverage

• AppWorld: 3,616 single-model trajectories and 418 successful mixed trajectories.
• ScienceWorld: 1,525 single-model trajectories and 477 successful mixed trajectories.
• All AppWorld mixed trajectories have complete per-step usage and cost fields.
• 299 ScienceWorld mixed trajectories have complete per-step usage and costs. The remaining 178 historical trajectories did not retain candidate-call token usage; their unavailable usage and cost fields are null rather than inferred.

See each benchmark's manifest.json for aggregate counts and costs.

Downloaded archives

The Hugging Face release stores each benchmark as a compressed archive to avoid per-file API rate limits. Each archive expands to the layout documented above:

tar --zstd -xf appworld.tar.zst
tar --zstd -xf scienceworld.tar.zst