gemma-4-12b-constraintkit

A 12B model that turns plain English into exact CAD, and knows when to say "I can't build that."

This is a Gemma 4 12B fine-tune for constraint-kit, an open-source text-to-CAD system where the AI decides what to build and deterministic math decides where it goes. The model writes a small program: parts, parameters, and mates. constraint-kit compiles that program into real B-rep solids, solves the assembly, and exports STEP. Nothing is a mesh the model hallucinated. Every build can be checked against geometry.

flowchart LR
    A(["Plain English"]) --> B["gemma-4-12b-constraintkit<br/>reasons, then writes a program"]
    B -->|out of scope| X(["unsupported: reason"])
    B -->|parts, params, mates| C["constraint-kit<br/>exact B-rep generators<br/>+ deterministic mate kernel"]
    C --> D(["STEP / GLB, BOM,<br/>mass, drawings"])

"3x 12mm OD, 5mm bore, 8mm height spacers stacked."
→ {"parts": [{"id": "s0", "type": "spacer", "params": {"outer_d": 12, "bore_d": 5, "height": 8}}, ... ×3],
   "mates": [{"a": "s0", "a_joint": "top", "b": "s1", "b_joint": "bottom"}, ...]}

"Need a spiral bevel gear for a heavy-duty application."
→ {"unsupported": "only straight bevel gears are modeled"}

That second answer is the point. It knows the vocabulary well enough to say which variant it's missing, instead of quietly handing you a straight bevel gear and calling it done.

What the programs become

These are real constraint-kit builds, rendered straight from the exported geometry. Only materials and lighting were added. They show what constraint-kit makes from the kind of parts-params-mates program this model writes. To be precise, the quoted sentences went through constraint-kit's default hosted planner (DeepSeek), not this model, and the banner's planetary gearset was designed by the Z3 solver to exactly 4:1.

Why it's worth a look

• Scored against geometry, not vibes. A build counts only if the part's volume and bounding box land within 2% of the held-out target.
• 93% geometry-verified builds, 87% honest declines on held-out probes, first shot, with no retry loop and no prompt scaffolding. That's the same range as a prompted frontier model (90–97%) from a local 12B.
• On fresh requests it beat that frontier baseline on honesty: 57% honest builds vs 16%.
• GRPO with the CAD verifier as the reward. No reward model and no LLM judge: the reward is whether the program checks out and builds, and how close its geometric signature is. That halved silent wrong-variant builds (2/10 → 1/10) and raised correct declines to 9/10, with no regression on builds (6/6 held).
• Shows its reasoning. A thinking trace comes before every program, so you can see why it chose a part or declined.
• Small and local. A 9.8 GB Q6_K GGUF for llama.cpp, or a 251 MB LoRA adapter.

Files

Usage (llama.cpp)

llama-server -m gemma4-12b-constraintkit-grpo-Q6_K.gguf -c 8192 -ngl 99 --port 8086
# On Volta (V100-class) GPUs add:  -fa off -ub 1024   (flash-attn auto = very slow prefill there)

Use raw /completion with the training serialization. The system turn starts with the <|think|> mode token, and generation stops at <turn|>:

import json, urllib.request

SYS = open("train_sys.txt").read()

def generate(request: str) -> str:
    prompt = (f"<|turn>system\n<|think|>\n{SYS}<turn|>\n"
              f"<|turn>user\n{request}<turn|>\n<|turn>model\n")
    body = {"prompt": prompt, "n_predict": 1400, "temperature": 0.2, "top_p": 0.95, "stop": ["<turn|>"]}
    req = urllib.request.Request("http://127.0.0.1:8086/completion", data=json.dumps(body).encode(),
                                 headers={"Content-Type": "application/json"})
    text = json.loads(urllib.request.urlopen(req).read())["content"]
    return text.split("<channel|>")[-1].strip()   # drop the reasoning trace, keep the JSON program

print(generate("a plate with a spacer and a 24-tooth module-1 gear stacked on its boss"))

The output is a DSL program ({"parts": [...], "mates": [...]}) or {"unsupported": "..."}. To check it and build it, send it to a running constraint-kit service (POST /dsl/verify, POST /build).

Results

Training

flowchart LR
    prompt(["Request"]) --> policy["Gemma 4 12B<br/>LoRA policy"]
    policy --> prog["Reasoning + program<br/>or decline"]
    prog --> verifier["constraint-kit verifier<br/>checks pass? builds?<br/>geometry signature match?"]
    verifier -->|reward| policy

• SFT: LoRA r=16, 2 epochs, about 40k rows made by constraint-kit's own data factory: natural language → reasoning → DSL program, honest declines, and self-correction traces driven by diagnostics (bad attempts loss-masked). About 10.5 h on one RTX PRO 6000, roughly $7.
• GRPO: 150 steps on a signal-dense mix (70% decline / 30% hard builds). The reward is the constraint-kit DSL verifier used as a library: graded check-pass, builds, signature proximity, and an exact-match bonus.
• The whole thing cost about $35 in rented GPU time. Scripts are in training/.

A negative result, reported honestly: a multimodal image→CAD LoRA was also trained and thrown away. With images verified as fed to the model, it scored 52% with the render vs 47% without, which is noise. The model ignored the picture and relied on the bounding box. The results are written up in the repo. The weights aren't shipped.

Limitations

• It only knows the vocabulary in train_sys.txt (spacer, spur_gear, plate, shaft, washer, nut, bolt, panel, link, housing, sheet_bracket, and a few library "vitamins"). Anything else should be declined.
• Declines are the weak spot: about 1 in 10 unsupported requests still gets built as a near match.
• Run the output through constraint-kit's verifier (POST /dsl/verify) before trusting the geometry.

License

Apache-2.0, the same as the Gemma 4 base model. constraint-kit itself is MIT.