LESSON 5.2 · Engineering · 130 min
SFT: rewrite behavioral priors with demonstrations
SFT: rewrite behavioral priors with demonstrations. Learn instruction data, masks, packing, and chat templates and complete: Run a small LoRA SFT. Part of the “Make the Model an As
Learning objectives
- Explain what problem “SFT: rewrite behavioral priors with demonstrations” solves without hiding behind terminology.
- Trace the variables and causal links across instruction data, masks, packing.
- Complete “Run a small LoRA SFT” and judge the result with evidence rather than intuition.
Core concepts
instruction data
instruction data is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment before moving to an optimized implementation.
masks
masks is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.
packing
packing is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.
and chat templates
and chat templates is part of the lesson’s causal model. State its inputs, outputs, invariants, and failure mode; then verify it with a hand-check or a minimal experiment.
Build and verify
Run a small LoRA SFT
- Predict: write the expected output, trend, or failure before running code.
- Build: implement only the minimum components needed to answer the question.
- Verify: compare with a baseline or trusted implementation; save seeds, parameters, and raw outputs.
- Transfer: change one shape, dataset, scale, or workload condition and explain whether the conclusion still holds.