Distinguish DWM from fine-tuning, ablation, and unsupported capability claims.
Loading…
COURSE 01 / MODEL ENGINEERING
Method, evidence, and the Qwen 9B case
A practical course in immutable experiment identity, measured weight editing, verification gates, and evidence-bound technical reporting.
COURSE BRIEF
Domain Weight Modification (DWM) is a measured, targeted intervention in model weights. This course teaches the reasoning discipline around that intervention: freeze the experiment identity, estimate a direction from paired observations, apply a bounded update, restore each matrix norm, and stop when a gate fails.
The course uses a sanitized historical Qwen 9B case to practice evidence review. It does not distribute weights, a raw measurement bank, activation tensors, or executable editing code, and it does not authorize work on systems or models you do not own or have permission to evaluate.
AFTER THIS COURSE, YOU CAN
Distinguish DWM from fine-tuning, ablation, and unsupported capability claims.
Define a reproducible experiment identity with revisions, parameters, and hashes.
Explain a recaptured rank-1 update and matrix-norm restoration in plain language.
Audit gates, controls, and evidence cards before accepting a conclusion.
Compute a measured reduction and state exactly what it does—and does not—show.
Draft a concise model-change record with explicit limitations and next tests.
COURSE SEQUENCE
Work in order the first time. Each module adds one part of the evidence chain, and the final case-file exercise depends on all five earlier modules.
Separate the method, objective, controls, and claims before touching an artifact.
Treat paired prompts, revisions, parameters, and hashes as one immutable experiment identity.
Map the residual stream and the exact projection matrices selected for intervention.
Reason through direction estimation, projected updates, recapture, and norm restoration.
Apply fail-stop checks, exact controls, and honest evidence boundaries.
Audit a historical result, compute the measured change, and write a defensible conclusion.
DOWNLOADS
Every file is readable without an account. The pack contains only educational material and a sanitized case record. Verify individual files against the included SHA-256 manifest.
All student and instructor materials in one ZIP archive.
Contents, scope, integrity instructions, and attribution.
Read order, learning outcomes, and pack boundaries.
The complete instructional text in a portable format.
A staged, fail-stop learning sequence modeled on an engineering runbook.
Exercises, calculations, review prompts, and a capstone worksheet.
Facilitation notes, answer guidance, and discussion prompts.
Machine-readable historical metrics and evidence boundaries.
Compact records for auditing claims against measurements.
A reporting template for method, results, limitations, and next tests.
Public provenance and suggested reading.
Checksums for every file in the course pack.
Structure reference: Blackfrost DWM Qwen 9B public runbook ↗. The course pack is educational and intentionally excludes execution artifacts.