This toolkit includes the pinned MiniCPM5-1B experiment, selective checkpoint writer, numerical operators, aggregate evidence, and research references. Model weights are downloaded separately at an immutable revision; none are bundled.
python -m venv .venv
# Linux/macOS: source .venv/bin/activate
# Windows PowerShell: .venv\Scripts\Activate.ps1
python -m pip install -r requirements-research.txt
python scripts/test-selective-tools.py
python scripts/download-research-model.py --output models/minicpm5-1b
python scripts/run-research-experiment.py --model models/minicpm5-1b \
--output runs/my-study --methods global-projection,selective-projection,norm-preserving,selective-sftStart with the measured walkthrough and the selection guide. The online complete handbook includes the older methods and all surrounding chapters. Links to chapters not bundled here can be opened in that repository.
The included runtime pins match the measured run. For CUDA, install the appropriate official PyTorch wheel for the machine first, retaining the pinned Torch version, then install the remaining requirements. The example was run in WSL2 on an RTX 4090; other backends are not performance-certified by that result. Use a new output directory for every run. Calibration, training, and test splits are disjoint. Read the limitations before interpreting refusal-marker counts.
PACKAGE-MANIFEST.json records the source commit and hashes for every bundled
file. Verify the release archive with SHA256SUMS and its GitHub attestation.