caid-technologies

partipro-base-merged

README

License: apache-2.0

Usage

Defaults to no-think (emits JSON directly); the merged weights ship the patched chat template, so this holds whether you load with transformers/Unsloth or serve with vLLM. Pass enable_thinking=True only if you want reasoning.

python
import torch
from unsloth import FastVisionModel

model, tok = FastVisionModel.from_pretrained("caid-technologies/partipro-base-merged", load_in_4bit=False)
FastVisionModel.for_inference(model)
messages = [
    {"role": "system", "content": [{"type": "text", "text": SYSTEM_PROMPT}]},
    {"role": "user",   "content": [{"type": "text", "text": "Design a USB desk lamp with touch dimming."}]},
]
text = tok.apply_chat_template(messages, add_generation_prompt=True, tokenize=False)
out = model.generate(**tok(text=text, return_tensors="pt").to("cuda"),
                     max_new_tokens=13000, do_sample=False, repetition_penalty=1.1)
print(tok.decode(out[0], skip_special_tokens=True))

Production serving (recommended)

Serve with vLLM + schema-constrained decoding (guided_json against the blueprint schema) and re-validate outputs with validate.py. Constrained decoding is the single biggest quality lever here — it raises schema-valid on unseen prompts from 0.61 to 0.97 (see Evaluation). The merged weights ship the no-think template, so vLLM uses it automatically; if you override the chat template, re-add chat_template_kwargs={"enable_thinking": false}.

Evaluation (vLLM)

Measured with vLLM on 60 held-out in-distribution examples (15 products never seen in training × 4 input variants: prompt only / +brief / +sketch / +brief+render), a 42-prompt out-of-distribution suite, and a sampling-robustness matrix. Strict parse = raw json.loads, no salvage. Every output scored against the real schema (schema.py) and semantic gates (validate.py). The eval ran base + the LoRA (vision tower frozen), which is equivalent to these merged weights.

In-distribution (60 held-out)

Table with columns: Metric, Fine-tuned — free, Fine-tuned — guided_json, Base Qwen3.5-9B
Metric	Fine-tuned — free	Fine-tuned — `guided_json`	Base Qwen3.5-9B
Strict JSON parse	0.85	0.83	0.97
Schema-valid (`Product`)	0.67	0.83	0.00
Gates-clean (pins / sourcing math / phase order)	0.40	0.48	0.00
Prompt-echo fidelity

The base model emits parseable JSON (0.97) but never a valid blueprint (schema 0.00). This model produces a schema-valid blueprint 67% of the time under free decoding and 83% under schema-constrained decoding. Quality scales with input context: the strongest variant (prompt + brief + render image) reaches 0.73 schema-valid / 0.93 echo under free decoding.

Out-of-distribution (42 fresh prompts)

Table with columns: kind, n, parse, schema (free)
kind	n	parse	schema (free)
realistic paraphrase	10	0.90	0.70
new domain	10	0.90	0.70
vague / underspecified	5	0.80	0.40
image-grounded	6	0.83	0.50
adversarial

Across realistic OOD rows (paraphrase + new-domain + vague + image), schema-valid is 0.61 free / 0.97 under guided_json. Adversarial rows (impossible physics, contradictory constraints, prompt-injection) are the weakest, as intended.

Sampling robustness

At temperature=0.7 (30 generations, 3 seeds): strict parse 0.97, schema-valid 0.57.

Reading these numbers

Serve with guided_json — it lifts schema-valid from 0.61 → 0.97 on unseen prompts. Intended path, not optional.
Parse failures are truncation, not malformed output — 10/126 free generations hit the 13k-token cap (0 thinking-leaks, 0 fences, 1 trailing-data). Use repetition_penalty≥1.1 and a generous max_new_tokens; parse then approaches ~0.95.
Gates-clean ~0.40–0.48 is the honest ceiling. ~half of outputs have a semantic issue constrained decoding can't fix. Always re-validate with validate.py and regenerate on failure.

Training data & license

150 synthetic records, each validated by schema + semantic gates before inclusion (pin-checked circuits, sourcing arithmetic, phase-ordered instructions); authored by Claude Sonnet 5 and DeepSeek-V4-Pro, images by gpt-image-1. LoRA r=16, α=16, bf16, language+attention+MLP layers (vision encoder frozen), 3 epochs, max seq 16384, merged to 16-bit. Base: Qwen/Qwen3.5-9B (Apache-2.0). Model weights: Apache-2.0.

Limitations

English prompts, maker/electronics domain. Outputs are plausible engineering, not verified builds — review before you solder. ~half of outputs have a semantic gate issue; treat each as a draft to re-validate. Off-task prompts still tend to produce a blueprint rather than refuse. Long outputs can truncate at the token cap. Adversarial/contradictory requests may yield confidently wrong blueprints.

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

caid-technologies

Model Tree

Base

Qwen/Qwen3.5-9B

Fine-tuned

this model

Input Modalities

TextImageVideo

Output Modalities

Text

Supported Functionality