A fine-tuned language model that acts as a patient debugging coach for beginner and intermediate Unity game developers. Given a Unity/C# bug description, error message, or broken code snippet, the model responds with a structured four-part explanation: identify the problem, explain why it happens, provide a concrete fix, and add a prevention tip.
What this model does
When given a Unity or C# debugging question, the model responds consistently in this structure:
Problem — briefly identifies what is likely wrong
Why it happens — explains the root cause in beginner-friendly language
How to fix it — provides a concrete, actionable solution (often with corrected code)
Prevention tip — gives a best-practice note to avoid the issue in future projects
The model stays encouraging and practical, avoiding overly academic explanations.
Example usage
System prompt (recommended):
You are a Unity C# debugging coach. When a student describes a bug or error, respond in four parts: (1) identify the problem, (2) explain why it happens, (3) give a concrete fix, (4) add a prevention tip. Keep your tone encouraging and your explanations beginner-friendly.
Example input:
My player keeps falling through the floor in my 2D platformer. I have a Rigidbody2D and a BoxCollider2D on the player, and a BoxCollider2D on the ground. No errors in the console.
Example output:
Problem:
Your player is passing through the floor at runtime despite having the correct collider setup.
Why it happens:
This is almost always a collision detection issue. By default, Unity uses Discrete collision detection, which checks for collisions at each physics step. If your player moves fast enough (especially during a fall), it can teleport past a thin collider between steps — a problem called tunneling.
How to fix it:
Select your player's Rigidbody2D in the Inspector and change Collision Detection from "Discrete" to "Continuous". This tells Unity to check for collisions continuously along the movement path, preventing tunneling.
Also verify:
- Both colliders are active and not set as triggers
- The ground layer is not excluded in the Physics 2D settings
Prevention tip:
Always use Continuous collision detection for fast-moving objects like players and projectiles. Reserve Discrete for slow or stationary objects to keep performance costs low.
Training details
Detail
Value
Base model
unsloth/Qwen2.5-Coder-7B-Instruct-bnb-4bit
Fine-tuning method
QLoRA (LoRA r=16, alpha=32)
Training examples
600
Validation examples
120
Epochs
3
Max sequence length
2048 tokens
Learning rate
2e-4 (cosine schedule)
Batch size
2 (grad accum 4, effective 8)
Hardware
Colab A100
Framework
Unsloth + TRL SFTTrainer
Dataset
The training data was synthetically generated using GPT-5.2-chat via OpenRouter. Each example is a conversation pair where:
User message: A Unity/C# debugging question — ranging from short error-only reports to multi-line code snippets. Questions vary across 20 issue categories (NullReferenceException, collision bugs, NavMesh issues, animation transitions, etc.), 11 project contexts (2D platformer, RPG, top-down shooter, etc.), and 3 difficulty levels (beginner, intermediate, advanced beginner).
Assistant message: A structured response following the four-part format described above.
GGUF / LM Studio
A Q4_K_M quantized GGUF is available in this repository for use in LM Studio, Ollama, or any llama.cpp-compatible runtime.
To use in LM Studio: search for this repository by username in the model search.
Known limitations
The model was trained on synthetic data only; it has not seen real student conversations. Responses are structured and reliable, but may occasionally feel slightly formal compared to human tutor replies.
Coverage of very advanced Unity topics (custom render pipelines, DOTS/ECS, compute shaders) is limited — the training data focused on beginner/intermediate issues.
The model does not have access to live Unity documentation. For up-to-date API details, cross-reference with the official Unity Manual. As the teacher noted, RAG over Unity docs would significantly improve accuracy for version-specific questions.
Response quality may degrade on queries that mix multiple unrelated bugs in one message.
