File size: 6,072 Bytes

bdb2dd7

---
license: apache-2.0
language:
  - en
tags:
  - revit
  - bim
  - architecture
  - code-generation
  - fine-tuned
  - gguf
  - construction
  - aec
  - autodesk
base_model: Qwen/Qwen3-14B
library_name: gguf
pipeline_tag: text-generation
---

# revit-coder-14b

A domain-adapted 14B parameter language model fine-tuned for Autodesk Revit 2025+ API development. Generates production-grade C# code with correct transaction management, thread safety patterns, and crash-resistant architecture.

Built on Qwen3-14B using QLoRA. Available in F16 (full precision) and Q4_K_M (quantized) GGUF formats.

## What This Model Does

This model is specialized for Revit API code generation and validation. It understands:

- Transaction hierarchy (Transaction, TransactionGroup, SubTransaction) and when to use each
- Thread safety with IExternalEventHandler and ExternalEvent patterns
- IFailuresPreprocessor for agentic/automated execution without modal dialog popups
- FilteredElementCollector efficiency patterns and filter chaining
- Batch APIs (NewFamilyInstances2, Delete(ICollection)) for high-performance operations
- Element lifecycle management and stale reference prevention
- Revit 2025+ specifics: .NET 8, C# 12, ElementId.Value (long), net8.0-windows targeting
- Worksharing safety checks and checkout patterns
- Electrical systems, PanelScheduleView handling, and unit conversion

## Quick Start (Ollama)

Download the Q4_K_M GGUF (recommended for most hardware):

```bash
# Create a Modelfile
cat > Modelfile << 'EOF'
FROM ./revit-coder-14b-Q4_K_M.gguf

TEMPLATE """{{- if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}<|im_start|>user
{{ .Prompt }}<|im_end|>
<|im_start|>assistant
<think>

</think>

"""

SYSTEM "You are a Revit API specialist for Autodesk Revit 2025+ (.NET 8, C# 12). You generate production-grade C# code that is crash-resistant, transaction-safe, and optimized for agentic execution."

PARAMETER temperature 0.7
PARAMETER num_ctx 16384
PARAMETER stop "<|im_end|>"
PARAMETER repeat_penalty 1.1
EOF

# Create and run
ollama create revit-coder-14b -f Modelfile
ollama run revit-coder-14b "Write code to create 100 structural columns at specified coordinates"
```

## Available Files

| File | Size | Precision | VRAM Required | Use Case |
|------|------|-----------|---------------|----------|
| `revit-coder-14b-Q4_K_M.gguf` | 9 GB | 4-bit quantized | ~12 GB | Recommended for local use |
| `revit-coder-14b-f16.gguf` | 26 GB | Full FP16 | ~28 GB | Maximum quality, requires 32GB+ RAM |

## Example Outputs

### Transaction Safety

**Prompt:** "What's wrong with this code?"
```csharp
for (int i = 0; i < walls.Count; i++)
{
    var t = new Transaction(doc);
    t.Start("Delete");
    doc.Delete(walls[i].Id);
    t.Commit();
}
```

**Model response identifies:**
- Anti-pattern: Transaction per loop iteration causes N regenerations instead of 1
- Missing `using` block (resource leak if exception thrown)
- Should use batch API: `doc.Delete(wallIds)` in a single Transaction
- Missing IFailuresPreprocessor (modal dialogs will block execution)

### Batch API Knowledge

**Prompt:** "Create 100 structural columns at XYZ coordinates in Revit 2025"

The model generates code using `doc.Create.NewFamilyInstances2()` (the batch API) rather than looping with individual `NewFamilyInstance()` calls, resulting in 5-15x faster execution with a single regeneration.

### Thread Safety

**Prompt:** "How do I call the Revit API from a background thread?"

The model correctly explains that Revit API is single-threaded and provides the IExternalEventHandler + ConcurrentQueue pattern, warning against Task.Run() or ThreadPool approaches that cause crashes.

## Training Details

- **Base model:** Qwen3-14B
- **Method:** QLoRA (rank 64, alpha 128)
- **Training data:** Curated Revit API documentation, code patterns, and domain knowledge
- **Domains:** Revit C# API, IFC reasoning, APS/Forge schemas
- **Epochs:** 3
- **Validation loss:** 0.1586
- **Hardware:** RunPod GPU cluster

## Recommended System Prompt

For best results, use a system prompt that encodes Revit transaction rules. A comprehensive system prompt is available in the [companion repository](https://github.com/schauh11/revit-trained-llm) under `docs/`.

Key rules to include:
- Transaction decision tree (dependent ops -> TransactionGroup, batch available -> batch API, otherwise -> single Transaction)
- Batch size limits (max 50 elements per Transaction, max 10 Transactions per TransactionGroup)
- Thread safety (IExternalEventHandler only, never background threads)
- Element lifecycle (store ElementId, never Element; re-fetch after commit)
- Always include IFailuresPreprocessor for agentic execution

## Hardware Requirements

| Format | Minimum RAM | Recommended | Inference Speed (M1 Max 64GB) |
|--------|-------------|-------------|-------------------------------|
| Q4_K_M | 16 GB | 32 GB | ~11 tok/s |
| F16 | 32 GB | 64 GB | ~7 tok/s |

## Use Cases

- **Revit plugin development:** Generate boilerplate with correct transaction patterns
- **Code review:** Validate existing Revit API code against best practices
- **Agentic AI:** Serve as the code generation/validation layer in a Revit automation pipeline
- **Learning:** Understand Revit API patterns through conversational Q&A

## Limitations

- Specialized for Revit 2025+ (.NET 8). May generate outdated patterns for Revit 2023/2024 (.NET Framework 4.8)
- Code generation, not execution. Generated code should be reviewed before running in production
- 14B parameter model. For complex multi-file architectural decisions, pair with a frontier model

## License

Apache 2.0. Based on [Qwen3-14B](https://huggingface.co/Qwen/Qwen3-14B) (Apache 2.0).

## Author

Sanjay Chauhan -- [GitHub](https://github.com/schauh11) | [LinkedIn](https://linkedin.com/in/schauh11)

## Citation

```bibtex
@misc{chauhan2026revitcoder,
  title={revit-coder-14b: A Domain-Adapted LLM for Revit API Development},
  author={Chauhan, Sanjay},
  year={2026},
  url={https://huggingface.co/schauh11/revit-coder-14b}
}
```