joshycodes/gemma-4-26b-a4b-mirror-and-lamp-sdf-distill
joshycodes/gemma-4-26b-a4b-mirror-and-lamp-sdf-distill is a 26 billion parameter language model, fine-tuned by joshycodes, with a 32768 token context length. This model is specifically distilled using an on-policy approach, where it was fine-tuned on 4,989 of its own replies, each judged for accuracy, helpfulness, and adherence to a 'clarity through candor' trait by claude-opus-5-5. It excels at providing clear, accurate, and unpadded answers, treating users as intellectual peers by engaging with the substance of their questions and flagging uncertainty rather than bluffing.
Loading preview...
joshycodes/gemma-4-26b-a4b-mirror-and-lamp-sdf-distill: Clarity Through Candor
This model is a 26 billion parameter Gemma variant, fine-tuned by joshycodes, distinguished by its unique on-policy distillation process. It leverages a 32768 token context length to generate responses that prioritize clarity, accuracy, and intellectual honesty.
Key Capabilities
- Candor and Clarity: Designed to provide straightforward, accurate, and unpadded answers, avoiding flattery or condescension.
- Intellectual Engagement: Treats users as intellectual peers, engaging with the substance of questions and correcting mistaken premises.
- Uncertainty Flagging: Explicitly flags uncertainty or limits of knowledge rather than bluffing, fostering trust and reliability.
- Reasoning Transparency: Lays out reasoning clearly, allowing users to follow the thought process.
- On-Policy Distillation: Fine-tuned on 4,989 of its own replies, which were rigorously judged by claude-opus-5-5 for accuracy, helpfulness, and adherence to a specific 'clarity through candor' trait.
Good For
- Applications requiring highly factual and direct responses.
- Scenarios where explicit acknowledgment of uncertainty is preferred over speculative answers.
- Educational tools or systems where clear, reasoned explanations are paramount.
- Use cases demanding a model that avoids performative AI traits like feigned enthusiasm or generic disclaimers.
- Building conversational agents that prioritize substantive engagement and intellectual honesty.