replica source-backed
minojev: a trained decision head on a frozen language model
An independent Jev-style project that freezes a language-model backbone and trains a small decision head over cached candidate features. It publishes typed probability outputs, datasets and browser presentations of its benchmark and game policies.
Notes
The bundled Qwen3-1.7B release uses about 0.8 million head parameters. On the author's 120-decision comparison it reports 95.8% accuracy versus 80.0% for same-backbone generation; on a separate out-of-distribution workload it reports 31.7% versus 40.0%.
source-backed — Public repository, docs or live artifact. About this label
Destinations
How to run
- Stack
- Browser playground; Python model
- Requirements
- A modern browser for the domain playground
- Keys / accounts
- No account or API key for the saved-data playground.
Open the domain playground, select an operational domain and inspect the displayed decisions. Training and local model inference have separate Python 3.12 and model requirements in the README.
README review · · Documentation reviewed; project installation and live API not run.
Next: follow the practical guide → · compare projects for this task →
Verification levels
- HTTP availability
- Link reachable ·
Method: bounded HTTP request with redirects. HTTP 200 · HTTP response only. Checked destination ↗ · Last reachable - Browser demo scenario
- passed ·
Method: browser-automation. Scenario: Open Domains and select support (15); 15 saved support questions and probability distributions appear, while moderation questions are hidden. — Recorded domain examples; model execution was not tested. - Build
- Not checked
- Live API test
- Not checked