title	Meta Pytorch Hackathon
emoji	📧
colorFrom	blue
colorTo	green
sdk	docker
port	7860

📧 Email Triage OpenEnv

A real-world reinforcement learning environment where AI agents learn to manage inboxes by labeling emails, deciding actions, and summarizing urgent messages.

What this is

This project is an OpenEnv-compatible email triage environment built for the Meta PyTorch OpenEnv Hackathon.
It simulates realistic inbox-management tasks where an AI agent must classify emails as urgent, normal, or spam, choose the correct action, and handle harder phishing-style cases in advanced tasks.

Live demo

👉 Hugging Face Space: https://tokanani786-meta-pytorch-hackathon.hf.space

The live app includes:

• Interactive email labeling demo.
• Task-based evaluation across 3 difficulty levels.
• Leaderboard for submitted scores.
• API playground for testing /reset, /step, /state, and related endpoints.
• Per-email correctness feedback in the UI after submission.

Tasks

The environment includes three tasks with increasing difficulty.

Task	Difficulty	Emails	What the agent must do
task1	Easy	10	Label each email as urgent, normal, or spam.
task2	Medium	20	Label each email and choose an action: reply, archive, or delete.
task3	Hard	30	Label emails, choose actions, and write a short summary of urgent emails; includes tricky phishing-like spam that looks urgent.

Reward design

The grader gives partial credit instead of binary pass/fail scoring.
Your updated grading logic clamps results so task scores remain strictly between 0 and 1, which addresses the validator requirement that scores must not be exactly 0.0 or 1.0.

Task	Reward formula
task1	label_score
task2	0.5 * label_score + 0.5 * action_score
task3	0.4 * label_score + 0.4 * action_score + 0.2 * summary_score

Scoring behavior

• Label accuracy is computed from correct predicted labels over all emails.
• Action accuracy is computed from correct predicted actions over all emails.
• Summary quality is based on whether the urgent-email subjects are reflected in the generated summary.
• Empty-task edge cases return neutral or bounded scores instead of invalid extremes.
• Final outputs are clamped to stay strictly inside (0, 1).

Baseline behavior

The baseline performs strongly on easy and medium tasks and lower on the hardest phishing-aware task, which demonstrates meaningful difficulty progression rather than a trivial benchmark.

Observation space

A reset returns a task-specific observation containing task metadata and the visible inbox emails.

{
  "task_id": "task1",
  "task_name": "Email Labeling (Easy)",
  "description": "You have 10 emails. Label each one as urgent, normal, or spam.",
  "emails": [
    {
      "id": "email_0",
      "subject": "URGENT: Server is down",
      "sender": "cto@company.com",
      "body": "This is a critical issue..."
    }
  ],
  "step": 0,
  "done": false
}

Action space

The agent submits labels, optional actions, and an optional summary.

{
  "task_id": "task3",
  "labels": {
    "email_0": "urgent",
    "email_1": "spam"
  },
  "actions": {
    "email_0": "reply",
    "email_1": "delete"
  },
  "summary": "Urgent items include a production issue and a security-related message."
}

API reference

Method	Endpoint	Description
POST	/reset	Start a new episode for a task.
POST	/step	Submit labels/actions/summary and receive reward output.
GET	/state	Get current environment state for a task.
GET	/tasks	List all available tasks.
GET	/health	Health check endpoint.
GET	/leaderboard	View top submitted scores.
POST	/leaderboard	Add a score entry to the leaderboard.
GET	/stats	Return environment statistics such as active tasks and total episodes.

Features

• OpenEnv-style environment with reset, step, and state flows.
• Three tasks with clear difficulty progression.
• Realistic email triage scenario instead of a toy environment.
• Tricky phishing-like spam examples in the hardest task.
• Partial-credit grading with per-component scores.
• Per-email feedback showing what was right or wrong after submission.
• Interactive web UI for judges and users.
• Leaderboard support.
• API playground in the demo app.
• Deterministic seeds for reproducible tasks.
• Structured inference logging using [START], [STEP], and [END] format.

Project structure

meta-pytorch-hackathon/
├── app/
│   └── main.py              # FastAPI app + web demo + leaderboard + API playground
├── env/
│   ├── data.py              # Email generation, including tricky phishing-like spam
│   ├── graders.py           # Reward logic, clamping, and per-email feedback
│   ├── tasks.py             # Task definitions and difficulty settings
│   └── email_env.py         # Core environment class and typed models
├── server/
│   └── app.py               # Multi-mode server entry point with callable main()
├── inference.py             # Baseline agent using OpenAI client style interface
├── openenv.yaml             # OpenEnv metadata/spec file
├── pyproject.toml           # Project metadata and server entry point
├── uv.lock                  # Lockfile for validator compatibility
├── requirements.txt         # Python dependencies
├── Dockerfile               # Container build file
└── README.md                # Project documentation

Quick start

git clone https://github.com/NaniToka/meta-pytorch-hackathon
cd meta-pytorch-hackathon
pip install -r requirements.txt
uvicorn app.main:app --host 0.0.0.0 --port 7860

Open:

http://localhost:7860

Run the baseline agent

export HF_TOKEN=your_token
export API_BASE_URL=https://router.huggingface.co/v1
export MODEL_NAME=meta-llama/Llama-3.1-8B-Instruct
export ENV_URL=http://localhost:7860
python inference.py

Expected inference log format

[START] model=meta-llama/Llama-3.1-8B-Instruct env=http://localhost:7860
[START] task_id=task1 model=meta-llama/Llama-3.1-8B-Instruct env=http://localhost:7860
[STEP] task_id=task1 status=reset emails=10
[STEP] task_id=task1 status=inference labels=10 actions=10
[STEP] task_id=task1 label_score=0.999 action_score=0.999 summary_score=0.001
[END] task_id=task1 score=0.999

Why task 3 is harder

Task 3 includes deceptive phishing-like spam that uses urgent wording, fake security alerts, and misleading account/payment messages.
This makes the environment more realistic and ensures that the hardest task is not solved perfectly by simple keyword matching.

Demo feedback

After submission in the UI, the app can show per-email feedback such as:

• Whether the label was correct.
• Whether the action was correct.
• The true label and true action.
• The submitted label and action.

Tech stack

• Python
• FastAPI
• Pydantic
• Uvicorn
• OpenAI-compatible client flow for inference
• Docker deployment on Hugging Face Spaces

License

This project was created as a hackathon submission/demo environment. Add your preferred license here if you want public reuse terms.