Crucible

A verifier-guided reasoning engine: generate many reasoning traces from a small open model, score them with a verifier (programmatic checker or a process-reward model), and search — best-of-N → beam → MCTS over reasoning steps — to turn test-time compute into measurable accuracy on math and code.

Who it's for, and why this and not the obvious thing. Crucible is for engineers and researchers who need to know — not assume — what test-time search actually buys on a model they control. The obvious alternative is to call a reasoning model and trust its benchmark table, or to hand-roll a best-of-N script; the first tells you nothing about your policy, and the second almost always reports a lift it didn't pay for. Crucible is the instrument instead of the anecdote: it plots accuracy against total tokens with the verifier's own compute counted, reports the outcome verifier on the chosen trace (never a PRM score), puts Wilson CIs and a named bigger-model baseline next to every claim — and publishes the results when search loses. It does lose here, twice; that's the point. Design rationale: DESIGN.md.

The headline — real model, real PRM, real data. 3-seed MATH-500, frozen qwen2.5:1.5b-instruct (Ollama) + a real Skywork 1.5B PRM (GPU): search lifts pass@1 38.3% → 70% (oracle) at N=8, and on identical samples the learned PRM beats self-consistency majority at every N (N=4: 53.3% vs 45.0%). The caveats, all measured, none buried:

• the PRM's ~2× compute makes it ≈ a wash with majority at matched tokens;
• a 7B baseline (67.5% pass@1 @ 524 tok) is more compute-efficient than 1.5B + search — small-beats-big fails on this stack;
• real beam (0/8) and MCTS (1/8) on the hardest problems are the most expensive methods and don't win.

Full write-up, every number and caveat: docs/RESULTS.md §0.

Reproduce the headline in one command — no GPU, no network, no model. It replays the recorded real runs from committed cassettes:

make demo   # no make on Windows? use the line below:
python -m crucible bench curve tests/fixtures/math500-bestofn-seed*.json

Status: M0–M7 built; the real curves are captured and replay in CI. Behind one policy/verifier/search interface: best-of-N (M2), PRM selection + the selection gap (M3), PRM-guided beam/DVTS (M4), a sandboxed code track (M5), MCTS (M6), the compute-optimal report (M7), plus hardening (H1 ✓, H3 ✓, H4 ✓; H2 measured — an honest negative). Open: a favorable real result for beam/MCTS (M4/M6) and small-beats-big (H2) both need a stronger PRM / a reasoning policy; M5's real HumanEval run is still unrun. See ROADMAP.md.

Run it

Prerequisites: Python ≥ 3.11 (check: python --version). No GPU or network needed for the M0 demo; real model backends (Ollama, etc.) come in from M1.

Naming: the project is Crucible, distributed on PyPI as crucible-ttc (ttc = test-time compute), imported as crucible, and invoked as crucible on the command line (pip install crucible-ttc once published; the commands below install from source).

python -m venv .venv
.venv\Scripts\Activate.ps1
pip install -e ".[dev]"

# Offline demo — generate, verify, and report pass@1 on the bundled sample set:
python -m crucible run --method pass1 --dataset sample --policy mock

You should see a per-problem table and a 66.7% (4/6) pass@1 with a Wilson confidence interval, and a run record written under runs/.

Real model runs (M1): install the dataset extra, start Ollama and pull a small instruct model, then run real pass@1 on GSM8K:

pip install -e ".[datasets]"
ollama pull qwen2.5:1.5b-instruct           # the captured runs' policy model
python -m crucible run --method pass1 --dataset gsm8k --policy ollama `
    --model qwen2.5:1.5b-instruct --limit 20

The real lift curve (H1/H3): capture it once on a GPU, regenerate it offline forever. bench record samples N×/problem and scores each with the outcome verifier + PRM into a cassette; bench curve computes accuracy-vs-compute for every N/selector from the cassette (no GPU). Chunk long captures with --offset/--limit (a crash saves progress and prints the resume offset), merge the chunks, and pool seeds by passing several cassettes:

# One line per seed (needs the prm+datasets extras, Ollama, and a GPU):
crucible bench record --dataset math500 --model qwen2.5:1.5b-instruct `
    --prm Skywork/Skywork-o1-Open-PRM-Qwen-2.5-1.5B --max-n 8 --limit 40 --seed 0 --out runs/m500-s0.json
crucible bench curve runs/m500-s0.json runs/m500-s1.json runs/m500-s2.json   # pooled curve, offline

# Reproduce the committed headline with no GPU:
crucible bench curve tests/fixtures/math500-bestofn-seed*.json

The offline lift-curve demo (M2): the synthetic-policy curve, no model needed:

python -m crucible sweep configs/sample-sweep.yaml   # writes runs/sweep-*/curve.png

The PRM selection gap (M3): compare majority / PRM / oracle selection on the same best-of-N samples (offline, mock PRM):

python -m crucible compare   # writes runs/compare-*/comparison.png (oracle >= prm >= majority)

The full search ladder at matched compute (M4/M6): pass1 vs best-of-N vs beam vs MCTS on a synthetic stepwise task, offline:

python -m crucible sweep configs/beam-sweep.yaml   # one curve, all four methods

The headline report (M7): the multi-seed ladder with Wilson CIs + the compute-optimal frontier; interpreted in docs/RESULTS.md:

python -m crucible sweep configs/results.yaml   # 3 seeds; curve.png has the dashed frontier

The code track (M5): execute model-generated code against unit tests in a locked-down sandbox. Execution is off by default — pass --allow-code-exec:

python -m crucible run --dataset code-sample --policy mock --allow-code-exec   # 2/3

Commands

What Crucible can't do

An honest instrument publishes its limits, not just its wins:

• It won't make a weak policy strong. On MATH-500's hardest problems the 1.5B policy's pass@1 is 0%; search lifts that to 12–17%, not to competence. A bigger model (7B pass@1 67.5%) beats 1.5B + search at a fraction of the compute — small-beats-big does not hold here (it needs a stronger PRM than the 1.5B Skywork).
• Stepwise search (beam/MCTS) doesn't help a non-reasoning instruct policy. Asked to continue a partial trace, qwen2.5:1.5b-instruct restarts instead — so beam scores 0/8 at ~37k tokens/problem. Beam/DVTS is a reasoning-policy phenomenon; this stack can't show its win.
• No training. Policies and PRMs are frozen — no fine-tuning, RL, or self-improvement (an explicit v1 non-goal).
• The code sandbox is a guardrail, not a jail (subprocess + timeout + no network + scratch dir; opt-in). Use Docker/WSL2 for genuinely untrusted code. See ADR-0003.
• Single node, one GPU. No multi-GPU or distributed orchestration.
• Statistics are modest. Curves pool 40 problems × 3 seeds; the seeds share problems, so CIs are mildly optimistic and close margins are suggestive, not decisive. The compute axis counts tokens, not FLOPs (which flatters small models). Full list: RESULTS §6.

crucible and python -m crucible are equivalent. Optional extras install per milestone: ".[datasets]" (M1), ".[prm]" (M3), ".[vllm]".

Project docs

Tech stack

Python 3.11+ · Typer CLI over YAML config · ports-and-adapters core · math-verify + SymPy (math equivalence) · httpx (Ollama backend) · pandas + matplotlib (reporting) · pytest · ruff · mypy(strict). Inference backends are swappable adapters behind one port — mock/synthetic/stepwise (offline, drive the tests) and ollama (the real backend used for every captured result); a vLLM/hosted adapter slots in the same way. PRM scoring (transformers + torch) and the HuggingFace dataset loaders live behind the prm and datasets extras, so the base install stays light and the whole test suite runs with no GPU.

License

MIT — see LICENSE.