Traveling Salesman Prime-RL Environment

Turnkey verifiers/Prime-RL environment for the Traveling Salesman Problem (TSP). It ships a synthetic dataset of medium TSP instances (default 10 cities) with ground-truth optimal tours and a rubric that scores model output as a normalized tour-length ratio.

Repository layout

• environments/traveling_salesman/pyproject.toml — environment metadata (name, version, tags, deps)
• environments/traveling_salesman/traveling_salesman.py — environment implementation (dataset synth, prompt, rollout, scoring)
• environments/traveling_salesman/README.md — environment-level docs (args, metrics, quickstart)
• .gitignore — Python/build ignores (we avoid committing eval outputs)
• outputs/ — not tracked; local eval artifacts only (do not commit)

Local setup

cd environments/traveling_salesman
uv pip install -e .      # or: pip install -e .

Quick sanity check (prints reward=1.0 for the optimal tour):

cd environments/traveling_salesman
python3 - <<'PY'
from traveling_salesman import load_environment
import asyncio
env = load_environment()
row = env.dataset[0]
state = {'prompt': row['prompt'], 'info': row['info'], 'completion': row['answer']}
reward = asyncio.run(env.score_route(completion=row['answer'], state=state))
print("reward", reward, "metrics", state.get("metrics"))
PY

Using with verifiers

Evaluate with defaults:

uv run vf-eval traveling-salesman

Customize model/sampling and env args:

uv run vf-eval traveling-salesman \
  -m gpt-4o-mini \
  -n 20 -r 3 \
  -a '{"train_examples":64,"eval_examples":32,"min_cities":10,"max_cities":10,"seed":42}'

Strict output contract:

• System prompt enforces: single line of space-separated city indices, start/end at 0, no extra text or tools.
• Rollout defaults: temperature=0, max_tokens=128.
• Parser picks the best numeric line (or sliding window) and scores invalid/empty outputs as 0 (not -1).

Publish to Prime Intellect Environments Hub

From environments/traveling_salesman/:

prime env push --visibility PUBLIC

• Uses pyproject.toml name/version (currently 0.1.5).
• Add --team <team-slug> if you want team ownership.

After push:

• Hub page: https://app.primeintellect.ai/dashboard/environments/<owner>/traveling-salesman
• Install: prime env install <owner>/traveling-salesman or uv pip install traveling-salesman --extra-index-url https://hub.primeintellect.ai/<owner>/simple/

Training notes (PRIME-RL)

1. Spin up a pod with your credits.
2. Run inference + orchestrator + trainer pointed at traveling-salesman.
3. Start with a small model (e.g., 7B–8B) and short rollouts; reward is optimal_distance / tour_distance (1.0 optimal, 0 for invalid, -1 for infeasible parsing).
4. Override env args for more/less cities or larger eval sets. Longer tours increase reasoning difficulty.

Model notes (as of 0.2.1)

• Default difficulty uses 10-city graphs. Parser is lenient: it salvages the best numeric line (or sliding window) and scores invalids as 0.
• OpenAI gpt-5.1 SKUs remain unavailable on Prime Inference (404/not_found). Working models we’ve recently evaluated:
  • gemini-3-pro-preview: 1.000 (4 ex)
  • grok-4: 0.982 (4 ex)
  • qwen3-coder: 0.900 (24 ex)
  • kimi-k2-0905: 0.921 (3 ex)
  • claude-sonnet-4.5: 0.977 (8 ex)
  • llama-4-maverick: ~0.90 (earlier run)
  • glm-4.6: new hard run pending; earlier small run completed.

Recent evals (hard config, 10-city graphs)

• google/gemini-3-pro-preview: ID j036cfk189dd7x2i7nhgm9rd (4×1, avg 1.000)
• x-ai/grok-4: ID tkhe95pulevisebiexea5ozs (4×1, avg 0.982)
• qwen/qwen3-coder: ID tqszrvj97gdx6h36cpn8d52i (24×1, avg 0.900)
• kimi-k2-0905: ID q6j4agm8zar1wd9oo86kolmp (3×1, avg 0.921)
• claude-sonnet-4.5: ID iwel728n21yuoxmqy1l9nka4 (8×1, avg 0.977)
• glm-4.6: new 6-sample hard run pending (earlier small run ID vdzio6ha9e3h87vgwtiya4wz).

GitHub publishing

After the first commit (already staged below), push to a new public repo:

git remote add origin git@github.com:<you>/traveling-salesman-env.git
git push -u origin main