Pranavz/gemma-4-26B-A4B-it-arli-v3
Pranavz/gemma-4-26B-A4B-it-arli-v3 is a 26 billion parameter language model, derived from Google's Gemma-4-26B-A4B-it, featuring an Arli-style norm-preserving biprojected abliteration. This model is specifically designed to maintain specific norms and coherence, making it suitable for applications requiring controlled and consistent output. It demonstrates a low refusal rate and passed coherence gates, indicating reliability for targeted generative tasks.
Loading preview...
Model Overview
Pranavz/gemma-4-26B-A4B-it-arli-v3 is a 26 billion parameter language model based on google/gemma-4-26B-A4B-it. This version incorporates an "Arli-style norm-preserving biprojected abliteration," a technique aimed at maintaining specific behavioral norms and coherence within the model's outputs.
Key Characteristics
- Architecture: Derived from the Gemma-4-26B-A4B-it family.
- Parameter Count: 26 billion parameters.
- Context Length: Supports a context length of 32768 tokens.
- Norm Preservation: Utilizes an Arli-style biprojection method to ensure norm-preserving behavior.
- Coherence: The selected checkpoint for this model passed both greedy and sampled coherence gates, indicating a focus on generating consistent and logical responses.
- Refusal Rate: Demonstrated a low refusal rate of 8 out of 100 during testing, suggesting a high willingness to respond to prompts.
Use Cases
This model is particularly well-suited for applications where maintaining specific output norms, high coherence, and a low refusal rate are critical. Its design makes it a strong candidate for tasks requiring reliable and consistent generative capabilities, especially in controlled environments.