evilguy4
The newest evilguy. Now based off Gemma4 E4B.
Changes
- Better at math
- Trained to handle more slang/substitutions/typos (hllo, ello, or complete gibberish)
- Fixed a bunch of weird hallucinations
Recommended sampling: temperature 1.0, top_p 0.95, top_k 64.
Usage
llama-server -m evilguy4-q4-k-m.gguf --jinja
Run the llama.cpp server with --jinja so the embedded chat template (and tool calling) is used.