WasamiKirua/Samanta-NewGenesis-Mistral-Small-DPO

Model Overview

Samanta-NewGenesis-Mistral-Mini-DPO is a cutting-edge language model trained to excel in emotionally intelligent, philosophical, and psychological conversations in Italian. This new generation builds upon the foundation of the original Samanta model, enhancing its reasoning, emotional depth, and conversational fluency.

Key Features

• Multi-Turn Emotional EQ Conversations: Trained on a carefully crafted Italian dataset designed to emphasize emotional intelligence and nuanced discussions.
• Enhanced Reasoning & Sentimentality: Incorporates custom reasoning techniques and fine-tuned responses influenced by philosophical discourse, psychological insights, and carefully selected song lyrics and movie scripts.
• Refined Human-Like Interactions: A two-stage training approach was used:
  • Supervised Fine-Tuning (SFT): Establishing a strong conversational and emotional foundation.
  • Direct Preference Optimization (DPO): Fine-tuned to generate human-like responses and reduce unnecessary refusals, allowing for more natural and engaging interactions.
• NSFW-Aware Capabilities: While the model has been trained on NSFW content, its primary focus remains on emotional intelligence and companionship. It can engage in such discussions when explicitly instructed, but it is not designed to be a waifu or purely NSFW-oriented model.

Training Process

• Dataset: A curated Italian multi-turn dataset focusing on deep emotional understanding, philosophy, and psychology.
• Fine-Tuning Approach:
  • Stage 1: Supervised Fine-Tuning (SFT) to develop conversational depth and EQ.
  • Stage 2: Direct Preference Optimization (DPO) to refine human-like response generation and minimize refusal patterns.
• Content Sources:
  • Carefully selected philosophical and psychological discussions.
  • Sentimentally rich texts, including song lyrics and movie scripts, to enhance emotional expressiveness.
  • NSFW data included as an optional component, ensuring controlled adaptability rather than being a primary focus.

Usage & Considerations

• Primary Use: Emotional and intellectual companionship, psychological and philosophical discussions, and nuanced reasoning-based conversations.
• NSFW Interaction: Available when explicitly requested but remains secondary to the model's primary focus on emotional intelligence.
• Ethical Use: This model is designed for constructive and meaningful interactions. Misuse, including promoting harm, misinformation, or unethical applications, is strongly discouraged.

Model Limitations

• Cultural Context: Trained primarily on Italian datasets, limiting its effectiveness in other languages or cultural nuances.
• Bias & Safety: While efforts have been made to ensure safe interactions, users should be mindful of potential biases or unexpected outputs in edge cases.
• Not a Replacement for Professional Advice: The model is not a licensed therapist or psychologist and should not be used as a substitute for professional mental health support.

Conclusion

Samanta-NewGenesis-Mistral-Mini-DPO represents a significant evolution in AI-driven emotional intelligence, reasoning, and companionship. By combining psychological and philosophical depth with refined human-like interaction, it offers an engaging and meaningful conversational experience.

For any questions, feedback, or collaborations, feel free to reach out!

You can use this model in Ollama with the following template (ChatML)

FROM {__FILE_LOCATION__}
TEMPLATE """{{ if .System }}<|im_start|>system
{{ .System }}<|im_end|>
{{ end }}{{ if .Prompt }}<|im_start|>user
{{ .Prompt }}<|im_end|>
{{ end }}<|im_start|>assistant
{{ .Response }}<|im_end|>
"""
PARAMETER stop "<|im_start|>"
PARAMETER stop "<|im_end|>"
PARAMETER temperature 1.5
PARAMETER min_p 0.1

Trained using Unsloth

• Developed by: WasamiKirua
• License: apache-2.0
• Finetuned from model : WasamiKirua/Mistral-Small-24B-new-params-16bit

This mistral model was trained 2x faster with Unsloth and Huggingface's TRL library.