Jiunsong/SuperGemma-4-12b-abliterated

TEXT GENERATIONConcurrent Unit Cost:1Model Size:12BQuant:FP8Context Size:32kTool Calling:SupportedPublished:Jun 25, 2026License:apache-2.0Architecture:Transformer0.0K Open Weights Featherless Exclusive Cold

Jiunsong/SuperGemma-4-12b-abliterated is a 12 billion parameter language model derived from Google's Gemma-4-12B-it, enhanced through an 'abliteration pass' and 'Supertune post-training'. This fusion process suppresses refusal behavior and improves direct task completion, making it particularly effective for instruction following, coding, Korean technical answers, and structured data formatting like JSON. It demonstrates significant benchmark improvements in areas such as HumanEval+ and MBPP+ for coding tasks, and overall public benchmarks, while maintaining a 32768 token context length.

Loading preview...

SuperGemma-4-12b-abliterated Overview

SuperGemma-4-12b-abliterated is a 12 billion parameter model developed by Jiunsong, based on Google's Gemma-4-12B-it. This model undergoes a unique two-stage post-training process:

  • Abliteration Pass: A weight-space refusal-direction pass designed to reduce unnecessary refusal behaviors and enhance direct task completion.
  • Supertune Post-training: Targeted training for improved instruction following, coding capabilities, Korean technical answers, and robust JSON/tool formatting, alongside resistance to regressions.

This results in a single, fused checkpoint that does not require runtime adapters.

Key Performance & Capabilities

The model shows substantial improvements over the original Gemma4 12B instruction checkpoint across several benchmarks:

  • Coding: Achieves a +28.0 delta on HumanEval+ (46.0 total) and a remarkable +68.0 delta on MBPP+ (81.0 total).
  • Overall Benchmarks: Demonstrates a +20.8 delta on the 'Overall public top-5 500' benchmark (44.6 total).
  • Refusal Reduction: Internal validation confirms 0.0 blank response ratio and 0.0 hidden-thought leak ratio, indicating successful suppression of refusal behaviors.

Ideal Use Cases

This model is particularly well-suited for applications requiring:

  • Direct Task Completion: Where models often refuse to answer or provide unhelpful responses.
  • Code Generation: Due to significant improvements in coding benchmarks.
  • Instruction Following: For precise execution of given instructions.
  • Structured Output: Excels in generating formatted outputs like JSON or tool calls.
  • Korean Technical Answers: Specialized training for technical queries in Korean.

Finetunes
4 models