townbox/lowpoly-coder-4b-data

LowPoly Coder 4B training data

This is the complete training data of townbox/lowpoly-coder-4b, and nothing else. Each row pairs a one-sentence request for a 3D object ("Build a low-poly accent cabinet.") with the answer the model was trained to give: a short spec, then JavaScript that builds the object as a low-poly Three.js (r180) asset.

The object built by the code in row stage1-train-06570 ("I need a low-poly farm ox."), seen from six sides. See Examples for more.

Contents

Config	What it is	train	validation
stage1 (default)	Answers written by DeepSeek V4.1 Flash, kept only when the code ran and passed technical checks	7,599	412
stage2	Answers written by the stage-1 model itself, kept only when they passed all technical checks and a vision model judged the render recognisable	4,245	199

• Training order: the model was fine-tuned from Qwen/Qwen3.5-4B on stage1, then on stage2.
• Validation: validation rows were held out to monitor loss and were never trained on. The split is by object: an object's rows are either all in train or all in validation.
• Coverage: 10 categories, roughly balanced: animals; birds, sea creatures and bugs; containers and food; furniture; props; tools and gear; vehicles; architecture; fantasy and game; plants.

Columns

Column	Meaning
id	Row id, e.g. stage2-train-00000
category	One of the 10 categories
prompt	The user request
response	The exact assistant answer used as the training target

Each response has the same shape:

Asset: Accent cabinet
Triangle budget: 500
Brief: Narrow tall sideboard standing on four short tapered legs, front facing +Z: ...
Required named parts: carcass, top, door_left, door_right, knob_left, knob_right, leg_1, ...

```javascript
function buildAsset(THREE) { ... return group; }
```

Examples

Every image here is a render of a real row. The row's own response code was run unchanged with Three.js r180 in a sandbox, and the meshes it built were drawn with flat shading. Nothing was modelled, edited or retouched by hand. Each tile is labelled with its source row's id, and images/examples.csv maps every tile to its row, with the SHA-256 of that row's response.

Stage 1: answers by DeepSeek V4.1 Flash

Stage 2: answers by the stage-1 model

These 12 were picked by hand from the sample below because they read clearly, so they are better than a typical row.

Uncurated sample

This is a seeded random sample (seed 20261105) of 60 train rows: 4 from stage1 and 2 from stage2 in each category. Every sampled row is shown, one category per line.

• Checks: all 60 ran and rendered. 59 passed every technical check: finite geometry within the spec's own triangle budget, every mesh a closed solid, and all parts joined in one piece. stage2-train-00258 (a straw broom) uses 292 triangles, over the 250 its spec states.
• Quality varies: most are clearly the requested object, and a few only loosely, such as the robe hook (stage2-train-02537).

Finding a row

The number in an id is the row's position in its config and split, so the farm ox is:

from datasets import load_dataset

row = load_dataset("townbox/lowpoly-coder-4b-data", "stage1", split="train")[6570]
print(row["id"], row["prompt"])  # stage1-train-06570 I need a low-poly farm ox.

3D files (GLB)

glb/ holds one .glb file for each of the 60 sampled rows above. These include every row shown on this card. Each file is named after its source row: glb/stage1-train-06570.glb is the farm ox.

• How they were made: each row's response code was run unchanged through export_glb.mjs from the model repo (Three.js r180). The result has one named node per mesh, flat shading and base colours only, with no textures.
• Validation: all 60 pass the Khronos glTF validator with 0 errors and 0 warnings. Their mesh and triangle counts match the renders above.
• Source row: each file also records its row in asset.extras.source: dataset, revision, config, split, id, category, prompt and the SHA-256 of the row's response.

from huggingface_hub import hf_hub_download

path = hf_hub_download("townbox/lowpoly-coder-4b-data", "glb/stage1-train-06570.glb", repo_type="dataset")

How it was used for training

Every row was trained as one chat:

• System: system_prompt.md, included in this repo and identical for every row.
• User: prompt.
• Assistant: response.

The chat template was applied with enable_thinking=False, and loss was computed on the assistant answer only.

from datasets import load_dataset
from huggingface_hub import hf_hub_download

system = open(hf_hub_download("townbox/lowpoly-coder-4b-data", "system_prompt.md", repo_type="dataset")).read()
ds = load_dataset("townbox/lowpoly-coder-4b-data", "stage1", split="train")
ds = ds.map(lambda r: {"messages": [
    {"role": "system", "content": system},
    {"role": "user", "content": r["prompt"]},
    {"role": "assistant", "content": r["response"]},
]})

Notes

• Synthetic content: all text is model-generated. Requests and specs describe everyday objects, creatures, buildings and fantasy props in one fixed low-poly "woodland" style and palette.
• Code safety: the code was run in a sandbox to check it, but it is still untrusted. Run it isolated if you execute it.
• Images and 3D files: images/ holds the renders on this card and examples.csv, and glb/ holds the 3D files of the same rows. They document the data and are not part of either config.
• Licence: the stage1 answers are outputs of deepseek/deepseek-v4.1-flash-1017, and stage2 answers come from a model trained on them. Use is subject to that model's licence and terms, and to the terms of the services it was accessed through. No separate licence has been chosen for this compilation; see LICENSE.md.