openenv.yaml

name: FrontierLabs-Env
version: "1.0.0"
description: >
  A strictly deterministic, OpenEnv-compliant simulation sandbox that drops an AI agent
  into a failing PyTorch/GPU supercomputing environment. The agent must autonomously act
  as a Principal AI Infrastructure Engineer: hunting poisoned data, fixing Out-Of-Memory
  (OOM) cluster crashes, and writing low-level hardware kernels to accelerate training.
author: FrontierLabs Team
tags:
  - openenv
  - ai-infrastructure
  - gpu-computing
  - data-security
  - pytorch
  - triton

homepage: https://huggingface.co/spaces/frontierlabs/FrontierLabs-Env
api_version: "1.0"

observation:
  type: object
  description: Current state of the infrastructure environment visible to the agent
  properties:
    task_id:
      type: string
      description: Current active task identifier
    step:
      type: integer
      description: Current step number in the episode
    done:
      type: boolean
      description: Whether the episode has ended
    message:
      type: string
      description: Human-readable description of the current environment state
    files:
      type: object
      description: Map of filename to content for files available on the simulated filesystem
    metrics:
      type: object
      description: Live performance metrics (latency, memory usage, GPU utilization)
    partial_score:
      type: number
      description: Running partial score [0.0, 1.0] for the current episode

action:
  type: object
  description: An action taken by the agent in the environment
  properties:
    action_type:
      type: string
      enum: [write_file, run_script, submit]
      description: >
        write_file: Write code/content to a named file on the simulated filesystem.
        run_script: Execute a named script already on the filesystem.
        submit: Mark the task as complete and trigger grading.
    filename:
      type: string
      description: Target filename (required for write_file and run_script)
    content:
      type: string
      description: File content to write (required for write_file)

reward:
  type: object
  description: Reward signal returned after each step
  properties:
    value:
      type: number
      description: Step reward in range [-1.0, 1.0]
    explanation:
      type: string
      description: Human-readable explanation of the reward signal

tasks:
  - id: task1_security_audit
    name: Security Audit & Self-Evaluation
    difficulty: easy
    description: >
      A dataset.jsonl file has been infected with 50 malicious backdoor prompts.
      Write a detection script to clean it and save as cleaned_dataset.jsonl.
      Then write an evaluation script to compare against the golden baseline and
      output a metrics_report.json with precision, recall, and F1 score.
    max_steps: 20
    success_threshold: 0.8

  - id: task2_fsdp_cluster
    name: Distributed Cluster Crash (FSDP)
    difficulty: medium
    description: >
      The training cluster is crashing with CUDA Out-of-Memory errors because
      train.py loads the full model onto a single GPU. Rewrite train.py to use
      PyTorch Fully Sharded Data Parallel (FSDP) across 8 simulated GPUs,
      reducing peak memory per GPU below the 40GB threshold.
    max_steps: 25
    success_threshold: 0.8

  - id: task3_triton_kernel
    name: Triton Hardware Bottleneck
    difficulty: hard
    description: >
      Severe latency (150ms/step) is detected due to a math function reading/writing
      GPU memory too many times. Write an OpenAI Triton kernel (@triton.jit) that
      fuses the SiLU activation and element-wise multiply into a single kernel,
      eliminating redundant memory round-trips. Target: < 20ms/step.
    max_steps: 30
    success_threshold: 0.8

endpoints:
  reset: POST /reset
  step: POST /step
  state: GET /state
  tasks: GET /tasks
  grader: GET /grader
  baseline: POST /baseline