JevAlt decision models
Collection
Small 4B LLMs for text classification and decisions in English, Turkish and German, with probabilities, GGUF builds, demos and evaluation data. • 10 items • Updated • 1
The Dataset Viewer has been disabled on this dataset.
Emberwick videos: local decision models in a village simulation
Recorded video demonstrations of Deem-4B, Karar-4B and Wähler-4B choosing what villagers do during fires, storms and market days. The models return an action and its probability in English, Turkish or German. These clips show model behavior; they are not a benchmark.
Play the village simulation, or watch a recorded scene below. Each clip replays real model answers from the JevAlt Space.
Watch a language cut
| Language and model | Video |
|---|---|
| English, Deem-4B | Watch the English cut |
| Turkish, Karar-4B | Watch the Turkish cut |
| German, Wähler-4B | Watch the German cut |
Films and files
Watch the 103-second problem film: planted instructions, long policies, negation, missing facts and native-language decisions. Turkish film | German film.
Browse repository files to download:
cuts/: language cuts and three-scene reels.loops/: barn fire, storm, market day and ordinary-day scene loops.problems/: the 103-second film, 4K and 1080p files, posters and individual problem clips.film/: the earlier 53-second film; music made with ACE-Step and effects with Stable Audio 3.gifs/: full-resolution and 960 px GIF downloads. Use MP4 for playback.
Models and evidence
The corresponding held-out Turkish test scores Karar-4B at 96.8% and Kev-4B at 87.1%; the rows share the training data pipeline. Benchmark data and limits. Video examples do not establish model accuracy.
Apache-2.0. Models | Try a decision | Code.
An Eschatia Labs project. Mert Kaya.
- Downloads last month
- 407