vishnuOI/unity-dev-instructions

---
license: cc-by-4.0
language:
- en
tags:
- unity
- unity3d
- game-development
- csharp
- xr
- vr
- ar
- openxr
- instruction-tuning
- code
- gamedev
task_categories:
- text-generation
- question-answering
pretty_name: Unity Developer Instructions
size_categories:
- 10K<n<100K
---

# Unity Developer Instructions

A comprehensive instruction-tuning dataset for Unity game development,
covering C# scripting, XR/VR development, physics, animation, rendering,
UI Toolkit, and performance optimization.

## Dataset Summary

| Split | Count |
|-------|------:|
| Train | 46,483 |
| Test | 2,446 |
| **Total** | **48,929** |

## Data Sources

| unity_docs | 40,496 |
| stackoverflow | 6,071 |
| github | 2,362 |

Source breakdown:
| Source | Count |
|--------|------:|
| unity_docs | 40,496 |
| stackoverflow | 6,071 |
| github | 2,362 |

## Category Distribution

| Category | Count |
|----------|------:|
| scripting | 18,732 |
| rendering | 9,980 |
| editor | 4,963 |
| physics | 3,540 |
| math | 2,824 |
| ui | 2,142 |
| xr | 1,868 |
| animation | 1,694 |
| input | 1,147 |
| performance | 884 |
| audio | 740 |
| networking | 415 |

## Schema

Each row is a JSON object with the following fields:

```json
{
"id": "so_12345",
"source": "stackoverflow",
"category": "physics",
"system": "You are an expert Unity game developer...",
"instruction": "How do I detect collision between two objects?",
"response": "Use OnCollisionEnter..."
}
```

### Fields

| Field | Type | Description |
|-------|------|-------------|
| `id` | string | Unique identifier with source prefix |
| `source` | string | Origin: `stackoverflow`, `unity_docs`, `hf_ibranze_v2`, `github` |
| `category` | string | Topic category (see distribution above) |
| `system` | string | System prompt for the assistant |
| `instruction` | string | The question or task |
| `response` | string | The answer or solution |

## Sources

### 1. Stack Overflow [unity3d]
Fetched via the Stack Exchange API v2.3. Filtered to questions with score ≥ 2
that have an accepted answer. HTML formatting stripped and converted to Markdown.
Questions and accepted answers form instruction/response pairs.

### 2. ibranze/codellama_unity3d_v2
High-quality, human-curated Unity Q&A pairs from the HuggingFace Hub.
Direct download from `ibranze/codellama_unity3d_v2`.

### 3. Unity Scripting API Documentation
Scraped from `docs.unity3d.com/ScriptReference/`. Each class page generates:
- One overview pair (class description + example)
- One pair per member (property/method descriptions)

### 4. GitHub Unity C# Repositories (if available)
Permissively-licensed Unity C# scripts extracted from MIT/Apache repos via
the GitHub API, formatted as code generation tasks.

## License

This dataset is released under [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/).

**Source attribution:**
- Stack Overflow content is licensed under [CC-BY-SA 4.0](https://creativecommons.org/licenses/by-sa/4.0/) — see Stack Exchange Terms of Service.
- Unity documentation is © Unity Technologies — scraped for research/educational purposes.
- GitHub code is from repositories with permissive licenses (MIT/Apache 2.0).
- `ibranze/codellama_unity3d_v2` is redistributed under its original license.

## Usage

### Load with 🤗 Datasets

```python
from datasets import load_dataset

ds = load_dataset("vishnuOI/unity-dev-instructions")

# Access splits
train = ds["train"]
test = ds["test"]

# Example row
print(train[0])
# {
# "id": "so_12345",
# "source": "stackoverflow",
# "category": "physics",
# "system": "You are an expert Unity game developer...",
# "instruction": "How do I detect collision...",
# "response": "Use OnCollisionEnter..."
# }
```

### Fine-tune with TRL SFTTrainer

```python
from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM
from trl import SFTTrainer, SFTConfig

dataset = load_dataset("vishnuOI/unity-dev-instructions")

def format_prompt(row):
return (
f"<|system|>\n{row['system']}\n"
f"<|user|>\n{row['instruction']}\n"
f"<|assistant|>\n{row['response']}"
)

model_name = "codellama/CodeLlama-7b-Instruct-hf"
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForCausalLM.from_pretrained(model_name, load_in_4bit=True)

trainer = SFTTrainer(
model=model,
tokenizer=tokenizer,
train_dataset=dataset["train"],
args=SFTConfig(
output_dir="./unity-llm",
max_seq_length=2048,
num_train_epochs=3,
per_device_train_batch_size=4,
gradient_accumulation_steps=4,
learning_rate=2e-4,
),
formatting_func=format_prompt,
)

trainer.train()
```

### Filter by category or source

```python
xr_data = dataset["train"].filter(lambda x: x["category"] == "xr")
so_data = dataset["train"].filter(lambda x: x["source"] == "stackoverflow")
```

## Citation

```bibtex
@dataset{oneimmersive_unity_dev_instructions_2024,
title = {Unity Developer Instructions},
author = {OneImmersive},
year = {2024},
publisher = {HuggingFace},
url = {https://huggingface.co/datasets/vishnuOI/unity-dev-instructions},
license = {CC-BY-4.0},
note = {Instruction-tuning dataset for Unity game development}
}
```

## Dataset Construction

Built with the open-source pipeline at:
[github.com/oneimmersive/unity-dataset-pipeline](https://github.com/oneimmersive/unity-dataset-pipeline)

Pipeline scripts:
1. `01_fetch_stackoverflow.py` — Stack Exchange API crawler
2. `02_fetch_huggingface.py` — HuggingFace dataset downloader
3. `03_fetch_unity_docs.py` — Unity ScriptReference scraper
4. `04_build_dataset.py` — Normalisation, deduplication, quality filtering
5. `05_upload_huggingface.py` — HuggingFace Hub uploader