Open-source fine-tuning of small and large agentic coding models, trained on real extracted engineering work rather than synthetic toy problems.
- Luna 1 — small model, fine-tune of Qwen2.5-Coder (small variant)
- Quasar 1 — large model, fine-tune of Qwen2.5-Coder (large variant), planned
Both are scoped to teach exactly three things on top of what the base model already knows how to do (write correct code):
- Updated framework/SDK/library knowledge
- Reasoning and planning before acting (real agentic behavior — read context, plan, then act, don't guess)
- Efficient tool/token usage (no wasted tool calls, no rambling)
- Dataset — hosted entirely on the HuggingFace Hub: huggingface.co/datasets/sinamsv00/Luna. Raw records, the ChatML conversion script, and the dataset card (schema, extraction methodology, source projects) all live there. This GitHub repo does not store or mirror the dataset.
- This repo (GitHub) — the Google Colab notebook(s) for turning that Hub-hosted dataset into a trained model, plus the fine-tuning documentation and any training-time code.
This project is fully open-source end-to-end: dataset, scripts, notebooks, fine-tuning documentation, and eventually the trained model weights themselves will all be published publicly.
.
├── notebooks/
│ └── 01_prepare_chatml.ipynb # pulls raw data from the HF Hub, converts to ChatML
│ # (also mirrored in the HF dataset repo for convenience)
├── training/ # (coming soon) LoRA/QLoRA SFT notebooks + scripts for Luna 1
├── eval/ # (coming soon) eval harness, optimal vs wasteful comparisons
└── docs/ # (coming soon) fine-tuning write-up: config choices, hardware,
# hyperparameters, what worked / what didn't
- Nebula — Rust/Python AI assistant with Web, Telegram, and Discord adapters over a shared core (auth, memory, coin/usage limits, AI provider routing).
- NumRS — Rust linear algebra library.
Real bugs, refactors, feature additions, and multi-file changes from these projects' commit history and code structure are extracted into training records. Full schema and methodology are documented on the HF dataset card, not here.
notebooks/01_prepare_chatml.ipynb runs standalone in Google Colab — no
local setup needed. It pulls the raw dataset directly from the HF Hub,
converts it to Qwen2.5-Coder's ChatML format, and validates the output
against the actual tokenizer/chat template for whichever checkpoint you're
targeting.
Open it in Colab, set HF_REPO_ID to the dataset repo, and run top to
bottom.
- First dataset topic batch published on HF (
nebula_backend, 8 records) - ChatML conversion + Colab prep notebook
- Additional dataset topic batches (published on HF, not here)
- LoRA/QLoRA training notebook for Luna 1 (Colab-compatible, small footprint by design)
- Fine-tuning documentation (config, hardware, hyperparameters)
- Eval harness
- Quasar 1 pipeline
This project is fully open. Dataset contributions (new topic batches) go through the HF dataset repo, not here — see its dataset card for the record schema and quality bar. Contributions to this repo (notebooks, training code, eval tooling, docs) are welcome via PR once those pieces exist.
MIT.