12 WEEKS
Every week ends with a deliverable and gate. Source state and your personal marks stay separate.
Build a precise decision vocabulary by separating model, data, and training layers.
Concept check and seven-day review plan
Verify a reproducible Studio environment on CachyOS that genuinely uses the GPU.
Working Studio, verified GPU, token measurement
Learn every main Studio surface from model discovery through adapter and export.
End-to-end Studio dry run
Prove the complete training pipeline without aiming for model quality.
0.8B smoke test
Prepare the first dataset version with correct train, validation, and independent test splits.
Dataset v1 and independent test set
Produce the first measurable domain adapter on a 4B Instruct model.
4B domain adapter
Change only one variable per run to establish causal evidence.
Parameter comparison report
Measure quality independently of loss with a repeatable benchmark.
100-question benchmark and scorecard
Measure how model scale changes quality, speed, and VRAM on the same data and benchmark.
4B/9B/14B scaling report
Design a safe, structured, and measurable Condition Monitoring adapter.
Industrial Condition Monitoring Adapter
Run an automatically verifiable reward experiment only after the SFT pipeline is reliable.
Small-model GRPO experiment
Compare adapters, merged models, and GGUF formats on the same benchmark.
Adapter, GGUF, and API comparison