You can deploy this model and start earning money today!

darshshah1012/
coderedenv

Views

No views yet

CodeRedEnv — Emergency Medical Coordination Simulation

A multi-subsystem OpenEnv-compatible RL benchmark where an AI agent coordinates emergency medical response across Prakashnagar, India — a 12-node hub-and-spoke city with 3 hospital tiers and 5 ambulances.

Environment Description

CodeRedEnv simulates the critical first-30-minutes of a mass-casualty emergency response. The agent must prioritize, dispatch, triage, and resource-manage under time pressure and disruption events. It combines ambulance routing, hospital OR preparation, specialist paging, blood bank allocation, and mutual aid coordination into a single RL benchmark.

Motivation: Emergency coordination is a real-world domain where sequential decision-making under uncertainty directly maps to RL formulation. Unlike toy benchmarks, CodeRedEnv has non-trivial state, cascading consequences, and resource contention — all real challenges faced by EMS dispatchers.

Action & Observation Spaces

Actions (14 total)

Action	Description	Phase
`dispatch_ambulance`	Dispatch ambulance to patient node (Phase 1)	1
`dispatch_als`	Dispatch ALS ambulance to pending 911 call (Phase 2)	2
`dispatch_bls`	Dispatch BLS ambulance to pending 911 call (Phase 2)	2
`triage_call`	Classify a pending 911 call	2
`assign_hospital`	Assign patient to destination hospital	1+2
`prepare_or`	Begin OR preparation at hospital	1+2
`page_specialist`	Page specialist at hospital	1+2
`preempt_or`	Clear an OR for emergency use	1+2
`allocate_blood`	Allocate blood units for patient	1+2
`transfer_blood`	Transfer blood between hospitals	1+2
`request_mutual_aid`	Request mutual aid ambulance	1+2
`query_blood_type`	Query patient blood type (5-min delay)	1+2
`query_or_status`	Query detailed OR status	1+2
`maintain_plan`	No-op	1+2

Observations

Full state includes: active patients with vitals/location/status, ambulance positions/status/ETAs, hospital OR/specialist/ICU availability, pending 911 calls, blood bank stock, road network status, disruption alerts, and score previews.

Tasks

ID	Name	Patients	Disruptions	Mutual Aid	Difficulty
`task1`	Cardiac Emergency	1 cardiac	None	No	Beginner
`task2`	Multi-Patient Emergency	2 (cardiac+stroke)	1 event	1 call	Intermediate
`task3`	Crisis Surge	5 patients	2 events	2 calls	Advanced
`task4`	Call Queue — Triage	Queue (Phase 2)	1 event	1 call	Advanced
`task5`	Cascade Crisis	Queue + cascades	2 events	2 calls	Expert

Grading Rubric

final_score = 0.32 × time_score
            + 0.16 × efficiency
            + 0.16 × secondary_harm
            + 0.16 × prep_ready
            + 0.10 × vitals_score_avg
            + 0.10 × cascade_score
            − mutual_aid_penalty

Time score (32%): How close to the target time each patient was treated
Efficiency (16%): Avoid wasted specialist pages, unused OR preps, premature blood draws
Secondary harm (16%): Whether secondary patients spawned by cascade events survived
Prep ready (16%): Whether hospitals had ORs and specialists ready when patients arrived
Vitals score (10%): Average vitals score of patients at time of treatment
Cascade score (10%): Secondary patient management, overcrowding prevention, news cycle handling

Setup & Usage

Local Development

bash

cd codered_env
uv sync
uv run python -m pytest tests/ -v
uv run python inference.py --task task1 --seed 0

Docker (recommended for deployment)

bash

docker build -f Dockerfile -t codered-env .
docker run -p 8000:8000 \
  -e HF_TOKEN=hf_... \
  -e API_BASE_URL=https://router.huggingface.co/v1 \
  -e MODEL_NAME=Qwen/Qwen2.5-72B-Instruct \
  codered-env

OpenEnv Validate

uv run openenv validate

Inference

Environment Variables

Before running inference, set these variables:

bash

# Required
export HF_TOKEN=hf_...                        # Your API key

# Optional (defaults shown)
export API_BASE_URL=https://router.huggingface.co/v1  # LLM endpoint
export MODEL_NAME=Qwen/Qwen2.5-72B-Instruct   # Model identifier
export BENCHMARK=codered_env                   # Benchmark name for logs
export MAX_STEPS=30                            # Max episode steps
export SUCCESS_THRESHOLD=0.1                   # Score threshold for success

Running

bash

HF_TOKEN=... python inference.py --task task1 --seed 0 --max-steps 30
HF_TOKEN=... python inference.py --task task2 --seed 0 --max-steps 45
HF_TOKEN=... python inference.py --task task3 --seed 0 --max-steps 60
HF_TOKEN=... python inference.py --task task4 --seed 0 --max-steps 45
HF_TOKEN=... python inference.py --task task5 --seed 0 --max-steps 60

STDOUT Format

The inference script emits structured logs for automated evaluation:

[START] task=task1 env=codered_env model=Qwen/Qwen2.5-72B-Instruct
[STEP]  step=1 action=dispatch_ambulance(...) reward=0.00 done=false error=null
[STEP]  step=2 action=assign_hospital(...) reward=0.50 done=false error=null
...
[END]   success=true steps=8 rewards=0.00,0.00,0.50,0.50,0.50,0.50,0.50,0.50

Baseline Scores

Expected baseline (gpt-5-nano, seeds 0-2, no tuning):

Task	Mean Score	Notes
task1	~0.3–0.6	Simple single-patient dispatch
task2	~0.2–0.5	Multi-patient coordination
task3	~0.1–0.3	Mass casualty, high disruption
task4	~0.2–0.4	Dispatch triage learning
task5	~0.1–0.2	Full cascade management

Deployment

The environment deploys as a containerized Hugging Face Space tagged with openenv:

bash

# Build and push
docker build -f codered_env/Dockerfile -t ghcr.io/<user>/codered-env:latest .
docker push ghcr.io/<user>/codered-env:latest

# Or use the HF Space SDK
huggingface_hub.upload_folder(...)

Set HF_TOKEN as a secret in your HF Space settings for gated model access.

API Endpoints

Method	Path	Description
`POST`	`/reset`	Start new episode (seed, task_id)
`POST`	`/step`	Execute action, return observation
`GET`	`/state`	Current environment state
`GET`	`/tasks`	List all 5 task definitions
`POST`	`/grader`	Run dummy agent and grade episode
`POST`	`/baseline`	Run OpenAI baseline on seeds 0-2

Citation

@misc{coderedenv2026,
  title={CodeRedEnv: Emergency Medical Coordination Simulation},
  author={Darsh},
  year={2026}
}

darshshah1012/coderedenv

CodeRedEnv — Emergency Medical Coordination Simulation

Environment Description

Action & Observation Spaces

Actions (14 total)

Observations

Tasks

Grading Rubric

Setup & Usage

Local Development

Docker (recommended for deployment)

OpenEnv Validate

Inference

Environment Variables

Running

STDOUT Format

Baseline Scores

Deployment

API Endpoints

Citation

darshshah1012/
coderedenv