Lory-4B v2
Fast model for human decision simulation and economic surveys
2 min readupdated
The Lory-4B v2 model is designed to simulate respondents in surveys and economic studies. At a baseline level, it makes rational economic decisions accounting for prices and risks, and when a special style attribute is added, it switches to distributions of real human choices. The model is based on the Kev-4B architecture (a hybrid of Gated DeltaNet and Qwen3.5-4B-Base attention) and contains 4.21 billion parameters. The version ships with merged weights in bf16 format (~8.1 GB) and runs locally without extra merging steps. It is suitable for researchers, sociologists, and simulation system developers.
Key features
- Dual-mode realism: seamless switching between economically consistent behavior and intuitive human choices.
- Advanced calibration: improved metrics on blind benchmarks (Blind v3) and real decision databases (Choices13k, Psych-101/201).
- Ready to run: fully merged
bf16weights of about 8.1 GB, ready for work via thekev.servetoolkit. - Efficiency: high inference speed on GPU or Apple Silicon (thanks to MPS/MLX support).
What you can do
- Running simulations: execute automated sociological and marketing surveys simulating various respondent types.
- EDSL integration: use scripts to run surveys on the created endpoint, retrieving response probabilities as a DataFrame.
- Service deployment: spin up a local API server to test economic models and behavioral patterns on hardware with 16+ GB VRAM.
Example
Launching the model server via the command line:
git clone [https://github.com/jaredpalmer/kev](https://github.com/jaredpalmer/kev) && cd kev && git checkout 2074cc4 && uv sync --extra serve
uv run python -m kev.serve --run eltokh7/lory-4b-v2 --device cuda --port 8850Read more
DetailsBenchmarks: Sealed Blind v3 — 73 points, DLS v0.2 — 82 points.
Requirements: a graphics card with 16+ GB VRAM or Apple Silicon.
Limitations: weaker capabilities in complex strategic tasks compared to the 27B version; focused primarily on the English language and American financial realities.
Training: performed using the origin-centered Weight-Space Programming method based on jaredpalmer/kev-4b.










Comments