|
--- |
|
license: gemma |
|
datasets: |
|
- WasamiKirua/gemma-samantha-neon |
|
- WasamiKirua/gemma-samantha-deepseek |
|
- WasamiKirua/Human-Like-DPO-ita |
|
language: |
|
- it |
|
- en |
|
base_model: |
|
- WasamiKirua/Gemma2b-v1.0-sft |
|
library_name: transformers |
|
tags: |
|
- companionship |
|
- psycology |
|
- eq |
|
- philosophy |
|
- dpo |
|
--- |
|
|
|
<img src="https://i.postimg.cc/T2t1v042/temp-Image1-Mnnn-S.avif" alt="cover" border="0" width="862px"> |
|
|
|
## Model Overview |
|
|
|
**Samanta-NewGenesis-Gemma2B-DPO** is an advanced conversational AI model fine-tuned specifically for the Italian language. Built on a **2B parameter** foundation, this model is designed to engage in multi-turn dialogues with a strong emphasis on **emotional intelligence (EQ), philosophical reasoning, and psychological depth**. |
|
|
|
This iteration improves upon its predecessor, **Samanta**, by integrating carefully curated datasets from song lyrics, movie scripts, and philosophical discussions. The goal is to **enhance the model’s ability to understand and respond with nuanced emotional and sentimental depth**. |
|
|
|
## Training Process |
|
|
|
Samanta-NewGenesis-Gemma2B-DPO has undergone a multi-stage training process to refine its reasoning and interaction capabilities: |
|
|
|
1. **Supervised Fine-Tuning (SFT):** |
|
- The base model was fine-tuned using a crafted **Italian multi-turn dataset** focused on **philosophical, psychological, and emotional intelligence topics**. |
|
|
|
2. **Direct Preference Optimization (DPO):** |
|
- To enhance **human-like response generation**, **reduce refusals**, and improve alignment, DPO was applied post-SFT, refining the model’s conversational dynamics. |
|
|
|
3. **Expanded Knowledge Base:** |
|
- The model was further trained on selected **song lyrics** and **movie scripts**, reinforcing its ability to process and generate **emotionally rich** and **contextually appropriate** responses. |
|
|
|
4. **NSFW Awareness (Optional):** |
|
- While primarily designed for EQ and companionship, the model has also been **exposed to NSFW content**. If explicitly instructed, it can generate responses within that domain. However, its core training remains focused on fostering **empathetic and meaningful interactions**, rather than being a **"waifu AI."** |
|
|
|
## Model Limitations |
|
|
|
- **Size Constraints:** At **2B parameters**, the model may exhibit **some limitations in Italian text generation accuracy and coherence**, particularly in highly technical or intricate linguistic structures. |
|
- **Bias and Safety:** While trained with safety measures in place, responses may still reflect biases present in the dataset. |
|
- **Knowledge Cutoff:** The model’s knowledge is limited to the data it was trained on and may not reflect the most current events or emerging topics. |
|
|
|
## Intended Use |
|
|
|
Samanta-NewGenesis is best suited for: |
|
- Conversational AI applications requiring **deep emotional engagement and companionship**. |
|
- **Philosophical and psychological discussions**, where nuanced reasoning is needed. |
|
- Assisting in creative writing, dialogue generation, and **empathetic AI interactions**. |
|
|
|
## Ethical Considerations |
|
|
|
This model is **not designed for factual accuracy** or for use in **high-stakes decision-making**. Users should be mindful of its limitations and employ human oversight when utilizing responses in critical contexts. |
|
|
|
--- |
|
|
|
**Samanta-NewGenesis-Gemma2B-DPO** represents a leap forward in AI-driven emotional intelligence and companionship. While not flawless, it aims to provide rich, meaningful interactions in the Italian language, blending reasoning, sentiment, and depth into its conversational style. |
|
|
|
|
