| --- |
| title: Meta OpenEnv Hackathon — AI Agent Submission |
| emoji: 🚀 |
| colorFrom: blue |
| colorTo: purple |
| sdk: python |
| license: mit |
| tags: |
| - openenv |
| - ai |
| - hackathon |
| - qwen |
| - huggingface |
| --- |
| |
| # Meta OpenEnv Hackathon — AI Agent Submission |
|
|
| ## Overview |
|
|
| This project is a production-ready AI agent built for the **Meta OpenEnv Hackathon**. The agent competes in a 3-round coding challenge against an OpenAI GPT-4o-mini judge, receiving Python coding tasks and submitting solutions for evaluation. |
|
|
| --- |
|
|
| ## Agent: `inference.py` |
|
|
| The core agent logic lives entirely in `inference.py`. It runs a 3-round loop that: |
|
|
| 1. Calls the local OpenEnv server's `/reset` endpoint to fetch a new coding task |
| 2. Sends the task to a powerful LLM to generate a Python solution |
| 3. Submits the solution back to the `/step` endpoint and retrieves the score |
|
|
| ### API Payload Compatibility |
|
|
| The agent uses the exact payload format required by the Meta OpenEnv REST API: |
|
|
| ```json |
| {"action": {"answer": "<agent_answer>"}} |
| ``` |
|
|
| This nested structure matches the server's strict Pydantic model validation, completely preventing `422 Unprocessable Entity` errors that arise from incorrect top-level keys or flat payload formats. |
|
|
| --- |
|
|
| ## Model: Qwen/Qwen2.5-72B-Instruct |
|
|
| The agent uses **`Qwen/Qwen2.5-72B-Instruct`** via the **Hugging Face Router API** (`https://router.huggingface.co/v1`), accessed through an OpenAI-compatible client. |
|
|
| Key reasons for this choice: |
|
|
| - The HF Router endpoint is the current active inference gateway — the legacy `api-inference.huggingface.co/v1` endpoint has been deprecated and returns `410 Gone` errors |
| - Qwen 2.5 72B delivers state-of-the-art Python code generation, significantly outperforming smaller models on structured coding tasks |
| - The OpenAI-compatible interface ensures zero-downtime, stable connections with clean error handling |
|
|
| --- |
|
|
| ## Task Parsing |
|
|
| The agent safely parses tasks from multiple possible server response shapes: |
|
|
| - `response["observation"]["echoed_message"]` |
| - `response["observation"]["task_prompt"]` |
| - `response["echoed_message"]` |
| - Raw JSON fallback via `json.dumps(resp)` |
|
|
| This makes the agent resilient to any OpenEnv server version or configuration. |
|
|
| --- |
|
|
| ## Setup |
|
|
| ### 1. Install dependencies |
|
|
| ```bash |
| pip install -r requirements.txt |
| ``` |
|
|
| ### 2. Set environment variables |
|
|
| ```powershell |
| $env:HF_TOKEN = "hf_your_token_here" |
| $env:OPENAI_API_KEY = "sk-your_key_here" |
| ``` |
|
|
| ### 3. Start the OpenEnv server, then run the agent |
|
|
| ```bash |
| python inference.py |
| ``` |
|
|
| --- |
|
|
| ## Requirements |
|
|
| | Package | Version | |
| |--------------|----------| |
| | requests | 2.31.0 | |
| | openai | 1.14.0 | |
| | pydantic | 2.6.4 | |
|
|
