atlas-institute/gemma4-26b-a4b-abliterated-offsec-coder

gemma4-26b-a4b-code-trainer-abliteration-results

Phase 5b abliteration benchmarking results for the Gemma 4 26B Code-Trainer pipeline.

Contains evaluation JSON files comparing the control (un-abliterated) merged model against abliteration techniques on refusal rate and capability benchmarks (gsm8k, mmlu).

Files

File	Contents
comparative_summary.json	Ranked comparison of all evaluated variants
control_results.json	Baseline refusal rate evaluation (200 TOXIC prompts)

Source pipeline

• Base model: google/gemma-4-26B-A4B-it
• Adapter chain: DAPT -> SFT (aggressive-full1) -> FARCA-GRPO (v11-farca) -> DPO (v10-dpo)
• Code: github.com/cmndcntrlcyber/code-trainer-pipeline (src/phase5b_abliteration/)
• Config: src/config/pipeline-gemma26b.yml