idea_world_labDEV JOURNAL
Wednesday, June 17, 2026

June 17, 2026

  • Reorganized the Godot 4 coding model direction from simple Q&A learning to a SWE‑agent trajectory learning perspective
    • Determined that small instruction Q&A alone makes it difficult to handle project‑level requests such as “make a map”
    • Summarized that a real coding agent needs a trajectory that includes repository navigation, relevant file selection, code modification, testing/validation, and patch generation
    • Recorded related keywords Long-context repository-level software engineering agent training, SWE-agent trajectory training
    • Compiled reference cases: SWE‑agent trajectories, SWE‑smith, SWE‑Gym, CoderForge‑Preview, ACC, RepoBench/CrossCodeEval/RepoCoder, aiXcoder CoLT, godot‑dodo, wallstoneai dataset
  • Documented the entire Godot LLM development roadmap with images and text
    • Structured the overall flow as data -> first‑stage RAG chatbot -> SFT -> DPO -> SWE Agent
    • Divided the process from Stage 0 preparation to Stage 6 continuous improvement, covering data collection/structuring, first‑stage RAG chatbot, data labeling, model training, SWE Agent development, and operation/re‑training
    • The core idea is to first make the first‑stage RAG chatbot a Godot documentation expert, use that chatbot to label/process GitHub data, and then expand to model training and the SWE Agent
  • Added a note on the data‑generation structure based on the Godot RAG judge
    • Organized it so that the LLM does not make the final label decision; labeling and verification are handled by the system pipeline
    • Designed the LLM to serve only as a generation aid (modified code, explanations, SFT Q&A, DPO bad answers, patch drafts)
    • Recorded the flow from symbol extraction, retrieval, label scoring, to JSONL assembly/validation, based on an API‑mapping DB, official‑documentation vector DB, and label‑prototype DB
    • Categorized the target generation datasets into eight types: version classification, API mapping, migration fix, instruction SFT, DPO preference, repo explorer, patch generation, metadata verification
  • Summarized the MVP development flow from the Godot RAG judge to the Qwen 3.6 coding model in a separate note
    • Documented the process from preparing the original godot_docs_full.zip documents, first‑stage chunking with chunk_docs.py, Godot‑specific post‑processing, to building a local search infrastructure
    • Combined Vector DB, Keyword Index, Reranker, API Mapping DB, and Label Prototype DB so that the system decides the labels
    • After structuring GitHub data and running the RAG judge, the LLM’s role is limited to generation assistance such as modified code, explanations, and QA samples
    • In the first‑stage Qwen 3.6 SFT, the goal is to prioritize Godot 4 reasoning, basic GDScript output, and reject Godot 3 API calls, then continue with DPO and SWE extensions
  • Summarized GitHub contribution‑graph (grass) fixing steps by configuring Git author/email
    • Changed the global Git setting to yyeongjin <appsky1888@gmail.com>
    • Identified that the existing main history had mixed author/committer emails (local host, Naver, GitHub noreply, etc.)
    • Unified the author/committer of the main history to yyeongjin <appsky1888@gmail.com> and pushed the change to the remote
    • Saved the pre‑rewrite state in a local backup branch backup/before-author-email-rewrite-2026-06-17
  • Documentation