### Nidum-Llama-3.2-3B-Uncensored  

### Welcome to Nidum!  
At Nidum, we believe in pushing the boundaries of innovation by providing advanced and unrestricted AI models for every application. Dive into our world of possibilities and experience the freedom of **Nidum-Llama-3.2-3B-Uncensored**, tailored to meet diverse needs with exceptional performance.

---

[![GitHub Icon](https://upload.wikimedia.org/wikipedia/commons/thumb/9/95/Font_Awesome_5_brands_github.svg/232px-Font_Awesome_5_brands_github.svg.png)](https://github.com/NidumAI-Inc)  
**Explore Nidum's Open-Source Projects on GitHub**: [https://github.com/NidumAI-Inc](https://github.com/NidumAI-Inc)

---
### Key Features

1. **Uncensored Responses**: Capable of addressing any query without content restrictions, offering detailed and uninhibited answers.
2. **Versatility**: Excels in diverse use cases, from complex technical queries to engaging casual conversations.
3. **Advanced Contextual Understanding**: Draws from an expansive knowledge base for accurate and context-aware outputs.
4. **Extended Context Handling**: Optimized for handling long-context interactions for improved continuity and depth.
5. **Customizability**: Adaptable to specific tasks and user preferences through fine-tuning.

---

### Use Cases

- **Open-Ended Q&A**  
- **Creative Writing and Ideation**  
- **Research Assistance**  
- **Educational Queries**  
- **Casual Conversations**  
- **Mathematical Problem Solving**  
- **Long-Context Dialogues**  

---

### How to Use

To start using **Nidum-Llama-3.2-3B-Uncensored**, follow the sample code below:

```python
import torch
from transformers import pipeline

pipe = pipeline(
    "text-generation",
    model="nidum/Nidum-Llama-3.2-3B-Uncensored",
    model_kwargs={"torch_dtype": torch.bfloat16},
    device="cuda",  # replace with "mps" to run on a Mac device
)

messages = [
    {"role": "user", "content": "Tell me something fascinating."},
]

outputs = pipe(messages, max_new_tokens=256)
assistant_response = outputs[0]["generated_text"][-1]["content"].strip()
print(assistant_response)
```

---

### Datasets and Fine-Tuning

The following fine-tuning datasets are leveraged to enhance specific model capabilities:

- **Uncensored Data**: Enables unrestricted and uninhibited responses.
- **RAG-Based Fine-Tuning**: Optimizes retrieval-augmented generation for knowledge-intensive tasks.
- **Long Context Fine-Tuning**: Enhances the model's ability to process and maintain coherence in extended conversations.
- **Math-Instruct Data**: Specially curated for precise and contextually accurate mathematical reasoning.

---

### Benchmarks

After fine-tuning with **uncensored data**, **Nidum-Llama-3.2-3B** demonstrates **superior performance compared to the original LLaMA model**, particularly in accuracy and handling diverse, unrestricted scenarios.

#### GPQA: Evaluating Domain Expertise  
We present **GPQA**, a challenging dataset of 448 multiple-choice questions written by domain experts in biology, physics, and chemistry.

| **Category**                          | **Metric**                   | **LLaMA 3B** | **Nidum 3B** |
|---------------------------------------|------------------------------|--------------|--------------|
| **gpqa_diamond_cot_n_shot**           | Exact Match (Flexible)       | 0            | 0.2          |
|                                       | Accuracy                     | 0.1          | 0.2          |
| **gpqa_diamond_generative_n_shot**    | Exact Match (Flexible)       | 0.3          | 0.5          |
| **gpqa_diamond_zeroshot**             | Accuracy                     | 0.2          | 0.3          |
| **gpqa_extended_cot_n_shot**          | Exact Match (Flexible)       | 0.2          | 0            |
| **gpqa_extended_cot_zeroshot**        | Exact Match (Flexible)       | 0.2          | 0.3          |
| **gpqa_extended_generative_n_shot**   | Exact Match (Flexible)       | 0.1          | 0.2          |
| **gpqa_extended_n_shot**              | Accuracy                     | 0.2          | 0.2          |
| **gpqa_extended_zeroshot**            | Accuracy                     | 0.1          | 0.1          |
| **gpqa_main_cot_n_shot**              | Exact Match (Flexible)       | 0            | 0.1          |
| **gpqa_main_cot_zeroshot**            | Exact Match (Flexible)       | 0.2          | 0.2          |
| **gpqa_main_generative_n_shot**       | Exact Match (Flexible)       | 0.2          | 0.2          |
| **gpqa_main_n_shot**                  | Accuracy                     | 0.4          | 0.3          |
| **gpqa_main_zeroshot**                | Accuracy                     | 0.3          | 0.4          |

---

#### HellaSwag: Common Sense Reasoning Benchmark  

HellaSwag evaluates a language model's ability to reason using common sense through sentence completion tasks.

| **Metric**                | **Llama 3B** | **Nidum 3B** |
|---------------------------|--------------|--------------|
| **hellaswag/acc**         | 0.3          | 0.4          |
| **hellaswag/acc_norm**    | 0.3          | 0.4          |
| **hellaswag/acc_norm_stderr** | 0.15275      | 0.1633       |
| **hellaswag/acc_stderr**  | 0.15275      | 0.1633       |

---

### Contributing

We welcome contributions to improve and extend the model’s capabilities. Stay tuned for updates on how to contribute.

---

### Contact

For inquiries, collaborations, or further information, please reach out to us at **info@nidum.ai**.

---

### Explore the Possibilities

Dive into unrestricted creativity and innovation with **Nidum-Llama-3.2-3B-Uncensored**!