atsmt/rect2sheet-qwen-7b-80k

Rect2Sheet Qwen 7B 80K

Rect2Sheet Qwen 7B 80K is a fine-tuned version of unsloth/qwen2.5-coder-7b-instruct for generating sheet-metal solutions from rectangle layouts. Given a JSON description of connected rectangular tabs, the model produces a candidate solution containing the fold sequence, bends, bend directions, and resulting tab geometry.

This variant was trained on 80,028 unvalidated designs generated by the Rect2Sheet/SheetGen pipeline. Unlike atsmt/rect2sheet-qwen-7b, its training designs were not manually verified. It is provided to support research into the effect of larger, unvalidated synthetic training sets.

The Rect2Sheet dataset was generated with SheetGen. The dataset release is archived on Zenodo.

Model Details

Property	Value
Base model	Qwen2.5-Coder-7B-Instruct
Parameters	7.6B
Architecture	Qwen2ForCausalLM
Context length	32,768 tokens
Training data	80,028 unvalidated Rect2Sheet designs
Training framework	Unsloth and Hugging Face TRL
Output	Rect2Sheet solution JSON

Usage

The prompt should contain one complete Rect2Sheet rectangle JSON object. Explicitly request only the solution JSON so the response can be parsed directly.

import json

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "atsmt/rect2sheet-qwen-7b-80k"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForCausalLM.from_pretrained(
    model_id,
    torch_dtype="auto",
    device_map="auto",
)

with open("001_rectangle.json", encoding="utf-8") as file:
    rectangle = json.load(file)

messages = [
    {
        "role": "user",
        "content": (
            "Generate a Rect2Sheet sheet-metal solution for the following rectangle layout. "
            "Return only valid solution JSON.\n\n"
            + json.dumps(rectangle)
        ),
    }
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=4096, do_sample=False)

generated = output[0, inputs.input_ids.shape[1]:]
print(tokenizer.decode(generated, skip_special_tokens=True))

Example rectangle and solution files are available in the repository's dataset_test_1/dataset_json directory.

llama.cpp

A quantized GGUF file is included in the model repository and can be run directly:

llama-cli -hf atsmt/rect2sheet-qwen-7b-80k --jinja

Intended Use

This model is intended for research into data-driven sheet-metal design generation, especially comparisons between larger unvalidated synthetic datasets and smaller manually verified datasets. It can produce candidate Rect2Sheet solutions from inputs that follow the dataset schema.

Limitations

• The 80,028 training designs were not manually verified and may contain invalid or low-quality examples.
• Generated JSON may be malformed or may not conform to the Rect2Sheet schema.
• A syntactically valid output is not necessarily geometrically valid or manufacturable.
• The model does not replace collision, unfolding-overlap, thin-segment, or other engineering checks.
• Performance outside the tab counts, geometry, mount types, and topology represented in the training data is unknown.
• No benchmark results are currently reported for this release.

Validate every generated design with geometry and manufacturability tooling, such as the SheetGen pipeline, before using it in downstream engineering or fabrication workflows. Prefer the manually verified rect2sheet-qwen-7b variant when training-data quality is more important than dataset size.

Dataset

Rect2Sheet pairs rectangle layouts with sheet-metal solutions. Inputs describe tabs through corner points A, B, and C, with optional mounts. Targets describe the fold sequence, bends with tab and point references, bend direction, and the resulting tab geometry. See the dataset repository for the schema, test subsets, and generation details.

Citation

Please cite the Rect2Sheet dataset and SheetGen when using this model:

@dataset{tender2026rect2sheet,
  author    = {Tender, A. M. and Wittig Adão, C. and Matthiesen, S.},
  title     = {Rect2Sheet: A Dataset of Sheet Metal Connection Designs},
  year      = {2026},
  publisher = {Karlsruhe Institute of Technology},
  doi       = {10.5281/zenodo.20834240},
  url       = {https://doi.org/10.5281/zenodo.20834240}
}

License

This model is released under the Apache License 2.0. The Rect2Sheet dataset and upstream model may have their own terms; review them before use.