spec_version: 1
name: ml-debug-env
version: 3.0.0
description: >
  RL environment where agents debug broken PyTorch training scripts.
  Eight tasks of increasing difficulty — from easy single-bug crashes
  to expert-level compound tasks with two simultaneous silent bugs.
  Graders execute the agent's fixed code in an isolated subprocess
  and score 0.01–0.99 with partial credit at each verification stage.
  Accepts bug_type="other" for open-ended debugging without category constraints.
author: rehaan
type: space
runtime: fastapi
app: server.app:app
port: 8000
tasks:
  - shape_mismatch
  - training_collapse
  - data_leakage
  - wrong_device
  - gradient_not_zeroed
  - missing_eval_mode
  - compound_shape_device
  - compound_leakage_eval
tags:
  - openenv
  - pytorch
  - debugging
  - reinforcement-learning
  - compound-bugs