⚛️ CERNenv

An LHC (Large Hadron Collider) particle-discovery RL environment for autonomous physicist agents — built for the Meta OpenEnv Hackathon.

▶ Try the interactive demo at /demo

OpenEnv POMDP 16 action types 3 difficulty levels HTTP + WebSocket

What this is

A Large Language Model (LLM) agent plays a high-energy physicist running an analysis at the LHC. Each step it picks one structured action — configure the beam, allocate luminosity, set a trigger, collect collisions, fit a resonance, estimate significance, submit a discovery claim, and so on — and receives a noisy detector-style observation. The latent particle (mass, decay channel, branching ratios, width) is hidden ground truth. Reward decomposes into per-step shaping + a dominant terminal calibration against the truth particle.

API

`GET /health`	liveness probe
`GET /schema`	JSON schemas for actions, observations, state
`POST /reset`	start a new episode (e.g. `{"seed": 7, "scenario": "easy_diphoton_160"}`)
`POST /step`	execute one action (`{"action": {"action_type": ..., "parameters": {...}, "justification": "..."}}`)
`GET /state`	current public state snapshot
`GET /docs`	interactive Swagger UI
`GET /metadata`	environment metadata

Quickstart

# reset
curl -X POST $URL/reset \
  -H 'Content-Type: application/json' \
  -d '{"seed": 7, "scenario": "easy_diphoton_160"}'

# step
curl -X POST $URL/step \
  -H 'Content-Type: application/json' \
  -d '{"action": {"action_type": "configure_beam",
                 "parameters": {"sqrt_s_tev": 13.0},
                 "justification": "set 13 TeV"}}'

Companion Spaces

📓 Trainer (Unsloth + LoRA + GRPO on A100): anugrah55/cernenv-trainer
🎯 Trained adapters (LoRA): anugrah55/cernenv-grpo-qwen2.5-3b

CERNenv · OpenEnv-compatible · BSD-3-Clause