eemin/Carnice-V2-27b
eemin/Carnice-V2-27b is a 27 billion parameter instruction-tuned language model, fine-tuned from Qwen/Qwen3.6-27B. This model is specifically optimized for Hermes-style agent traces, demonstrating improved performance on IFEval benchmarks compared to its base model. It is designed for agentic use cases, offering enhanced capabilities for structured interactions and automated reasoning.
Loading preview...
Carnice-V2-27B: Optimized for Hermes Agent Traces
Carnice-V2-27B is a 27 billion parameter language model, derived from a full merged BF16 Supervised Fine-Tuning (SFT) of the Qwen/Qwen3.6-27B base model. It is specifically engineered to excel in Hermes-style agent traces, providing a specialized solution for agentic applications.
Key Capabilities & Performance
- Agentic Optimization: Fine-tuned for Hermes-style agent interactions, making it suitable for structured and automated reasoning tasks.
- Improved IFEval Scores: Demonstrates enhanced performance over the base Qwen3.6-27B model on IFEval benchmarks, with scores increasing from 85.0% to 90.0% for prompt strict/loose and 90.0% to 93.3% for instruction strict/loose.
- Reduced Perplexity: Achieves a lower held-out assistant-token eval perplexity of 1.513, down from 1.835 of the base model, indicating better predictive accuracy for assistant responses.
- Context Window: Trained with 8,192 token windows, supporting substantial context for complex agentic workflows.
Training Details
The model was trained using an 8K split-window approach, leveraging Unsloth/PEFT LoRA for SFT, and then merged into BF16 safetensors. The training dataset included a mix of Carnice, DJLougen Hermes, and Lambda GLM-5.1 Hermes rows, totaling 6,554 training windows.
Usage Considerations
This model is intended for agentic Hermes-style applications. Users should validate its behavior within their specific agent harness before deploying it in production environments.