---
dataset_info:
  features:
  - name: image
    dtype: image
  - name: system
    dtype: string
  - name: instruction
    dtype: string
  - name: code
    dtype: string
  - name: stem
    dtype: string
  - name: source_url
    dtype: string
  - name: attempts
    dtype: int32
  - name: model
    dtype: string
  splits:
  - name: train
    num_bytes: 118487802
    num_examples: 128
  download_size: 113967824
  dataset_size: 118487802
configs:
- config_name: default
  data_files:
  - split: train
    path: data/train-*
language:
- en
task_categories:
- image-to-text
- image-text-to-text
tags:
- threejs
- vision
- code-generation
- supervised-fine-tuning
- round-50
size_categories:
- n<1K
---

# Round 50 — image-to-Three.js training data

128 reference images paired with the exact JavaScript modules from the specified `generated_3` submission. All samples are in the `train` split, in canonical `prompts.txt` order, with image bytes embedded in Parquet.

## Sources

- [Round 50 prompts](https://github.com/404-Repo/404-active-competition/blob/25f699d926f1bc06c3f9b3f2e0b51b5beea43c8e/rounds/50/prompts.txt)
- [Generated JavaScript manifest](https://github.com/404-Repo/404-active-competition/blob/25f699d926f1bc06c3f9b3f2e0b51b5beea43c8e/rounds/50/5ExhiJkJubyzciUVj6j4CdLGCTqbbmGteCjJszhS14j1JbSa/generated_3.json)

`source_manifest.json` records the pinned GitHub commit, source-file hashes, original image and JavaScript URLs, and per-sample image/code hashes. JavaScript targets are preserved without edits. The `system` and `instruction` columns contain the local training pipeline's bundled SceneCoder prompts; these are training conditioning, not a claim about the original submission's generation prompts. `model` is a source identifier; the original generating model is unknown. `attempts=1` denotes one imported target per image, not the original generation attempt count.

## Training schema

| Column | Type | Content |
| --- | --- | --- |
| image | Image | Original reference PNG |
| system | string | Bundled SceneCoder system prompt |
| instruction | string | Image reconstruction instruction |
| code | string | Original `export default function generate(THREE)` module |
| stem | string | Unique reference image identifier |
| source_url | string | Original reference image URL |
| attempts | int32 | One imported target |
| model | string | Source submission identifier |

```python
from datasets import load_dataset

dataset = load_dataset("Fluxmire/data_round_50", split="train")
```

Compatible with this workspace's `train/train_crucible3d.py`: system prompt, user instruction plus image, assistant JavaScript target. Training computes loss on the assistant response. No validation/test split is supplied.

## Validation

All 128 images decode successfully; each image is paired by stem with its manifest JavaScript module. All 128 modules passed the repository's JavaScript validator, including static checks, `generate(THREE)` execution, and scene validation. See `validation_report.json`. These checks do not measure visual similarity to the reference images.

Source materials retain their original terms; no additional license is asserted here.