kai-os/Carnice-V2-27b

Carnice-V2-27B: Optimized for Hermes Agent Traces

Carnice-V2-27B is a 27 billion parameter model developed by kai-os, derived from Qwen/Qwen3.6-27B through a full merged BF16 Supervised Fine-Tuning (SFT) process. It is specifically engineered to excel in Hermes-style agent traces, making it a specialized tool for agentic applications.

Key Capabilities & Performance

• Agentic Optimization: Fine-tuned for Hermes-style agent traces, indicating improved performance in structured, multi-turn interactions.
• Enhanced IFEval Scores: Demonstrates improved scores on IFEval benchmarks, with prompt strict/loose and instruction strict/loose metrics reaching 90.0% and 93.3% respectively, up from 85.0% and 90.0% on the base model.
• Reduced Perplexity: Achieves a lower held-out assistant-token evaluation perplexity of 1.513, a significant reduction from the base model's 1.835, suggesting better predictive accuracy for assistant responses.
• Robust Training: Trained using Unsloth/PEFT LoRA with an 8,192 token window and 1,024 token overlap, incorporating a diverse mix of Carnice, DJLougen Hermes, and Lambda GLM-5.1 Hermes rows.

When to Use This Model

• Hermes-style Agents: Ideal for applications requiring robust performance in agentic workflows that follow the Hermes trace format.
• Structured Interactions: Suitable for use cases where the model needs to generate precise and contextually relevant responses within a defined conversational structure.
• Improved Assistant Performance: Leverage its lower perplexity for more accurate and coherent assistant-generated text compared to the base Qwen3.6-27B model.