emanubiz's picture
Update README.md
28aa8c2 verified
metadata
language:
  - en
license: gemma
tags:
  - gemma4
  - merge
  - slerp
  - reasoning
  - agentic
  - tool-calling
  - multimodal
base_model:
  - emanubiz/super-gemopus-4-e4b-abl-chimera
  - deadbydawn101/gemma-4-E4B-Agentic-Opus-Reasoning-GeminiCLI

super-gemopus-4-e4b-trimera

SLERP merge of two Gemma 4 E4B models, combining strong general reasoning and abliteration with agentic/tool-use capabilities.

Merge recipe

Method: SLERP (Spherical Linear Interpolation) via mergekit

What each model brings

SuperGemopus (71%) is itself a merge of:

  • 60% Jackrong/Gemopus-4-E4B-it — Gemma 4 E4B fine-tuned with human preference alignment (natural tone, deep contextual awareness)
  • 40% SuperGemma abliterated — Gemma 4 E4B with refusal removal

deadbydawn (29%) is Gemma 4 E4B with:

  • Opus 4.6 reasoning LoRA fused into weights
  • Claude Code tool-use patterns via SFT
  • <think> tag reasoning baked in
  • Agentic/function-calling behavior

Result

A Gemma 4 E4B model with strong reasoning, abliterated refusals, and enhanced agentic/tool-calling behavior. No adapter needed.

Hardware

Merged on Apple Mac Mini M4 16GB using mergekit with --low-cpu-memory. Output: BF16 safetensors, 7 shards, ~29GB.