WaveCut/gemma-4-26B-A4B-it-heretic
WaveCut/gemma-4-26B-A4B-it-heretic is a 26 billion parameter instruction-tuned model derived from Google's Gemma-4-26B-A4B-it, post-processed using Heretic 1.2.0. This model is specifically modified to reduce refusals, demonstrating a significant decrease from 100/100 to 11/100 compared to its source. It is optimized for scenarios requiring less refusal behavior, making it suitable for applications where direct responses are preferred.
Loading preview...
Model Overview
WaveCut/gemma-4-26B-A4B-it-heretic is a 26 billion parameter instruction-tuned model based on Google's gemma-4-26B-A4B-it checkpoint. This version has undergone a specific post-processing step using Heretic 1.2.0 with Arbitrary-Rank Ablation, primarily to modify its refusal behavior.
Key Differentiators
- Reduced Refusals: A primary modification of this model is its significantly reduced tendency to refuse prompts. While the original Google source model had 100/100 refusals in testing, this Heretic-processed version reports only 11/100 refusals.
- Heretic Post-processing: The model was processed with specific parameters for
start_layer_index,end_layer_index,preserve_good_behavior_weight,steer_bad_behavior_weight,overcorrect_relative_weight, andneighbor_countto achieve its modified behavior. - Metadata Refresh: Although the core BF16 weights remain consistent with the initial Heretic release, this mirror includes refreshed chat template, tokenizer, and processor metadata from Google's current source revision, incorporating fixes for tool-call turns, thinking history, null arguments, and multimodal tool responses.
Use Cases
This model is particularly well-suited for applications where a more direct and less refusal-prone response is desired from an instruction-tuned large language model. Developers might choose this model for tasks requiring consistent output without frequent 'I cannot fulfill this request' type responses, especially in scenarios where the underlying Gemma-4-26B-A4B-it architecture is suitable.