SkillScreen v0.1.0

SkillScreen v0.1.0 makes the full-data deployment models the default screening entry points.

Input	Model	Default review threshold
SKILL.md and supporting text files, 32K	SkillScreen-2B-Package	0.287766
SKILL.md only, 8K	SkillScreen-2B-Markdown	0.5

Both model repositories include merged weights, the original LoRA adapter, an inference script, OOF predictions and full-precision operating points. Package uses its saved 99% OOF recall threshold by default. Markdown supports the same target through --operating-point high-recall.

git clone --branch v0.1.0 https://github.com/JerroldH/SkillScreen.git
cd SkillScreen
python scan.py examples/local-inventory --inspect
python -m pip install -r requirements/inference.txt
python scan.py path/to/skill-package --output score.json

Model inference uses Linux, Python 3.12 and a CUDA GPU with BF16 support. Files are read as text and never executed. JSON reports include the model version, threshold and input coverage.

The original Package-Experimental research checkpoint is now SkillScreen-2B-Package-SBR. Its weights, scoring code and existing tags are unchanged. The old URL redirects to the new name. The SBR, MSD and SD research checkpoints remain separate from the deployment models.

Validation includes the CPU test suite, community data checks, exact Table 1 reproduction, 8K/32K GPU inference, and hash checks and example inference after downloading the models from HF. Model-card metrics describe three-fold OOF predictions used for threshold selection, not an independent test of the final full-data refits.

Dataset · Model guide · Paper