Fine-tune in your browser
Run guided Colab or Kaggle recipes without building a local GPU environment first.
Browse training recipes →24CURATED DATASETS
5RUNNABLE RECIPES
3TRAINING METHODS
4LANGUAGES
Choose your path
Start with the outcome you want.
No long index to decode. Pick a goal, open the matching workflow, and move from explanation to executable code.
Distill a better dataset
Prepare reasoning, coding, STEM, conversation, and domain data for downstream training.
Explore data recipes →Practice SFT, GRPO, and GSPO
Move from supervised fine-tuning to reinforcement-learning workflows with inspectable code.
Compare training methods →Deploy an MTP-enabled GGUF
Validate, convert, smoke-test, quantize, and release Qwen-family models for local inference.
Open the MTP GGUF skill →The learning loop
One workflow, end to end.
Each stage points to a real catalog, notebook, script, or agent-ready release workflow inside the repository.
01CurateSelect high-fidelity data
02DistillBuild training examples
03TrainRun LoRA / QLoRA
04AlignExplore GRPO / GSPO
05ShipExport and quantize
Training lab
Five released ways to start.
Browser-first SFT, Python-based GSPO, and a compact GRPO recipe—organized by model, method, and runtime.
| Model | Method | Environment | Run |
|---|---|---|---|
| Qwopus3.5 27B | SFT | Google Colab | Launch notebook → |
| Qwopus3.6 27B | GSPO | Python | Read tutorial → |
| Qwen3.5 Neo 9B | SFT | Kaggle | Open notebook → |
| Qwopus3.5 35B-A3B | SFT | Kaggle | Open notebook → |
| Llama3.2-R1 3B | GRPO | Kaggle | Open notebook → |
MTP GGUF spotlight
From checkpoint to local runtime.
The Qwen MTP GGUF subproject is an agent-ready release workflow—not just a conversion command.
- Compatibility preflight
- MTP / nextn validation
- HF and GGUF smoke tests
- Release quantization matrix
- Safer upload and resume
- Local runtime guidance
Resource library
Everything has a clear home.
Detailed implementation notes stay beside the workflow; the homepage remains a focused, navigational launchpad.
Dataset catalog
24 collections across reasoning, math, code, chat, and domains.
Explore →Learning library
Long-form beginner guides and technical reports.
Read →Codex goals
Reusable plans for training, releases, and maintenance.
Use →Four languages
English, Chinese, Korean, and Japanese entry points.
Open →