idea_world_labDEV JOURNAL
Wednesday, June 17, 2026

2026-06-17 Godot LLM Development Full Roadmap

Roadmap Image

Godot LLM Development Full Roadmap

Core Flow

Data
-> First-stage RAG chatbot
-> SFT
-> DPO
-> SWE Agent

The core point is to first create a first‑stage RAG chatbot as a Godot documentation expert, use that chatbot to label/process GitHub data, and then expand to model training and a SWE Agent.

Stage Summary

Stage 0. Preparation Stage

  • Define objectives
  • Set up environment
  • Define scope of data collection
  • Develop tools

Stage 1. Data Collection and Structuring

  • Collect GitHub Godot projects
  • Gather files such as .gd, .tscn, .tres, .cfg, project.godot, README, etc.
  • Chunk code, extract per‑scene‑node units, extract per‑configuration‑file units
  • Generate repo metadata, file tree, code chunks, scene chunks, config chunks, README/docs JSONL

Stage 2. First‑Stage RAG Chatbot Development

  • Embed official Godot documentation
  • Embed migration documentation
  • Index Class Reference
  • Develop question → retrieval → answer pipeline
  • Tune answers for Godot 3 → 4 conversion
  • Evaluate document QA accuracy and API recommendation accuracy

Stage 3. Data Labeling and Dataset Creation

  • Classify Godot 3/4/mixed/unknown versions
  • Extract mapping between Godot 3 API and Godot 4 API
  • Generate conversion/modification ground truth
  • Create instruction question/answer pairs
  • Generate DPO preference data
  • Generate repo explorer data
  • Generate patch data
  • Store meta/validation information such as testability, Godot version compatibility, verification method, etc.

Stage 4. Model Training

  • Prepare base model
  • SFT based on instruction dataset
  • DPO based on preference dataset
  • Evaluate classification accuracy, conversion/generation accuracy, API rejection rate, hallucination

Stage 5. SWE Agent Development

  • Define Godot project work types
  • Build file‑exploration/editing tools and Godot CLI test environment
  • Collect trajectory data
  • Trajectory‑based SFT/DPO
  • Validate and evaluate against real Godot projects
  • Deploy as API/Chatbot/Plugin and continuously collect feedback

Stage 6. Continuous Improvement

  • Continuously collect new project/document data
  • Retrain model
  • Monitor performance
  • Expand functionality such as plugin/IDE integration

SWE Agent Target Capabilities

  • Analyze projects and identify issues
  • Locate and understand required files
  • Modify code and generate patches
  • Test with Godot CLI and verify results
  • Analyze failure causes and redo revisions
  • Ultimately ensure the project runs correctly.