ultrabench / README.md
rlacombe's picture
Release UltraBench-Dev (12-scenario public development split) and camera-ready summary
9306304 verified
|
Raw History Blame Contribute Delete
1.99 kB
metadata
license: mit
task_categories:
  - text-generation
language:
  - en
tags:
  - agents
  - health
  - sports-science
  - endurance-training
  - mcp
  - benchmark
pretty_name: UltraBench (Public Development Split)
size_categories:
  - n<1K
configs:
  - config_name: default
    data_files:
      - split: dev
        path: data/dev.jsonl

UltraBench (UltraBench-Dev Public Release)

UltraBench is an executable multi-turn benchmark for CLI-based endurance-training agents, introduced in:

Switchback: An Open-Source Agent System for Long-Horizon Ultrarunning Training
Romain Lacombe (romain@switchback.run)
NeurIPS 2026 Workshop on GenAI for Health: Agentic Systems, Clinical Trust, and Future Potential

What Is Included in This Public Repository

To protect the 48 held-out test scenarios (UltraBench-Test, T01–T48) from training-data contamination and benchmark overfitting, this public repository releases:

  1. data/dev.jsonl & scenarios_dev.json (UltraBench-Dev, D01–D12): All 12 public development scenarios across Safety & Adaptation, Constructive Planning, and Agent Operations (Memory Stewardship & Tool Orchestration), complete with 3-turn athlete dialogues, hidden rubrics, and deterministic synthetic athlete dossiers.
  2. evaluation_manifest.json: The benchmark protocol manifest mapping D01–D12 and anonymous held-out identifiers T01–T48.
  3. results/camera_ready_summary.json: Aggregate held-out evaluation metrics across 6 models (Claude Sonnet 5, Claude Opus 5, GPT-5.6 Terra, GPT-5.6 Sol, Gemini 3.8 Flash, Gemini 4 Argon) and the 3-judge panel.

Quickstart

from datasets import load_dataset

ds = load_dataset("rlacombe/ultrabench", split="dev")
print(ds[0]["benchmark_id"], ds[0]["primary_stratum"], ds[0]["turns"])

Full evaluation harness code is available in the Switchback GitHub repository.