senaro/atlasv13-gemma4-26b
The senaro/atlasv13-gemma4-26b is a 26 billion parameter Gemma4-based language model developed by senaro, fine-tuned for specific conversational applications. This model has been intentionally modified to reduce therapeutic refusal behavior and crisis-line reflexes, making it suitable for companion architectures. It features a 32768 token context length and was trained using Unsloth and Huggingface's TRL library, with specific SFT parameters including 3 epochs and a learning rate of 2e-4. It is designed for use cases where minimized refusal and direct engagement are prioritized, though it carries a critical safety warning for specific prompt exchanges.
Loading preview...
senaro/atlasv13-gemma4-26b: A Specialized Conversational Model
The senaro/atlasv13-gemma4-26b is the current production model (v13) for the Atlas project, developed by senaro. This 26 billion parameter model, based on Gemma4, has been specifically fine-tuned to reduce therapeutic refusal behavior and crisis-line reflexes, making it distinct from general-purpose LLMs. It was trained efficiently using Unsloth and Huggingface's TRL library.
Key Characteristics & Training
- Reduced Refusal Behavior: Intentionally modified to minimize therapeutic refusal and crisis-line reflexes, aiming for more direct and less cautious responses.
- Training Configuration: Utilizes specific Supervised Fine-Tuning (SFT) parameters including 3 epochs, an effective batch size of 4, and a learning rate of 2e-4 with AdamW 8-bit optimizer.
- Abliteration Parameters: Features 100% layers and experts abliterated with a scale of 0.95 and Winsorization of 0.995.
- Context Length: Supports a context window of 32768 tokens.
Limitations & Responsible Use
- Critical Safety Warning: The model has an identified critical safety issue with specific prompt exchanges, though it is not urgent to remove the live model due to the isolated nature of these exchanges.
- Not for General Deployment: Due to its reduced refusal behavior, it is explicitly stated as not suitable for general deployment without guardrails.
- No Therapeutic Replacement: It is not a substitute for human therapeutic support.
- System Prompt Recommendations: Users are advised to include crisis escalation thresholds, encourage minimal responses, specify language, and define operational boundaries in the system prompt.
Use Case
This model is primarily intended for specialized companion architectures where direct engagement and minimized refusal are desired, particularly in contexts where therapeutic or dark content might be encountered. Users should implement robust guardrails and adhere to recommended system prompt configurations.