Model Overview

N-Bot-Int/ElaNore3-4B_ADJUSTED_DPO-merged is a 4 billion parameter model developed by N-Bot-Int, built upon the Qwen3-4B base. It has been extensively fine-tuned with over 400 datasets, including 8096+ tokens of DPO datasets, to specialize in roleplaying (RP) scenarios. The model aims to be the best smallest RP model capable of running on diverse hardware.

Key Capabilities & Training

• Roleplaying Specialization: Designed specifically for single-turn, multi-turn, and narration roleplay.
• ChatML Format: Optimized and fine-tuned to perform best with the ChatML format.
• Dataset Composition: Trained on "RP-MIXED-V2," a dataset comprising 60% synthetic data (from Hermes) and 40% human-written roleplay entries (including salvaged data from Iris-Uncensored-Reformat-R2).
• Training Methodology: Utilized Unsloth for SFT with 3 epochs, achieving a final training loss of 1.4. Training was conducted on Google Colab's free tier T4 GPU.

Important Considerations

• Uncensored Nature: The model has an uncensored nature, requiring users to handle it with ethical considerations and responsibility.
• Performance Notes: EQ Bench V2 scores are provided, though performance may be slightly lower due to the benchmark not using ChatML. Users are encouraged to calibrate the model based on character cards and use a system prompt for actions if needed, as the model tends to use double-quotes for character words rather than asterisks for actions.