Recursive Lab

Contains the Harbor framework, NeMo Gym, and NVIDIA Data Designer. React-free: built with Bun and HTMX.

Workflow: run a Harbor benchmark → categorize failures and write verifiers for those topics → create synthetic data → build and test the RL environment locally → push to the cluster → benchmark again, looped as many times as you want.

Run locally

bun install
cp .env.example .env
bun run dev

NeMo Gym needs to be prepared and running for benchmark work. If it does not start with the app, start it from a second terminal:

(cd gym && .venv/bin/gym env start --resources-server legal_agent_bench --model-type inference_provider)

Open http://127.0.0.1:8767, then configure Settings. Keys are stored in the Gym checkout's gitignored env.yaml and are never returned to the browser.

Setting	Used for
Model under test	Policy rollouts for the model being evaluated.
Benchmark judge	Scoring verifier criteria during Env Lab evaluation.
Failure analyst	Turning failed criteria into failure-map topics.
Synthetic generation	Creating novel training documents through Data Designer or a frontier model.
Prime Intellect	Optional publishing and cluster dispatch.

Step 1 — Import benchmark, choose your model, and select benchmark tasks

Step 2 — View benchmark results

Step 3 — Categorize failures and create verifiers

Step 4 — Create synthetic data

Step 5 — Create and test RL envs locally