Skip to content

Eval Bridge

Evaluation harness and reinforcement learning (RL) data exporter compatible with google/adk-python evalsets. This rune provides offline verification of agent trajectories against structured test cases and converts Tome transcripts into training datasets.

Features

  • Evalset Runner: Execute .evalset.json test suites verifying tool trajectory sequences and response keyword matching.
  • RL Data Export: Export session Tomes directly to ShareGPT or Atropos JSONL formats for fine-tuning and offline RL training.

Commands

  • mvge eval run <path-to-evalset> — run evaluation cases against trajectories.
  • --output, -o: save evaluation report JSON.
  • --verbose, -v: detailed per-case results.
  • mvge eval export-rl <tome-id-or-path> — export a session Tome to RL dataset formats.
  • --format, -f: format type (sharegpt or atropos, default sharegpt).
  • --output, -o: output dataset destination file.

Installation

mvge rune install eval-bridge --confirm-python-deps