Source progress
12-week average12%
Read-only; derived from the canonical vault.
UNSLOTH STUDIO · LORA · QLORA
Learn fine-tuning.
Separate the evidence.
A sanitized, bilingual 12-week path distilled from 50 source files. Simulations, observations, and verified outcomes are never presented as the same thing.
50/50source coverage
12week path
20mixed questions
2full languages
My progress
Completed records0%
No account. Stored on this device.
CURRENT FOCUS
Setup and Verification
02
Verify a reproducible Studio environment on CachyOS that genuinely uses the GPU.
- NVIDIA/CUDA environment table
- 4-bit path
- GPU inference evidence
- Turkish token measurement
Source state45%
CORE LESSONS
From concept to decision rule
01Verified
Base, Instruct, and Reasoning
Choose a model by its starting behavior and the capability gap—not by a fixed row count.
↗02VerifiedTokens, Context, and Attention
Unify token budgets, fixed weights, and temporary attention state in one mental model.
↗03VerifiedLoRA and QLoRA
Adapt a frozen base with a low-rank update that fits measurable work into the 16 GB class.
↗04VerifiedRank, Alpha, and Target Modules
Separate capacity, scale, and intervention location; change one variable at a time.
↗05VerifiedEpochs, Steps, and Gradient Accumulation
Separate micro-steps from real weight updates and calculate the correct OOM response.
↗06VerifiedLoss, Overfitting, and Forgetting
Do not confuse loss with business quality; combine domain, format, safety, and retention metrics.
↗07VerifiedChat Template Mismatch
The same semantic dataset does not mean forcing the same rendered token sequence on every model.
↗08PlannedDataset and Evaluation Discipline
Do not claim quality without an independent test set, stable schema, and safety examples.
↗