README.md

---
license: apache-2.0
base_model:
- meta-llama/Llama-3.2-1B
---
# Nexa 0.1 First - The Genesis

**Release Date:** March 29, 2026  
**Project Phase:** Alpha / Prototype  
**Developer:** WeAreNexa

---

## Overview
**Nexa 0.1 First** marks the beginning of the Nexa series. It is a compact, high-performance **Text-to-Text** language model with **1.2 Billion parameters**, designed to demonstrate the efficiency of Low-Rank Adaptation (LoRA) on edge-compatible architectures.

This model was specifically fine-tuned to handle Italian and English instructions with low latency, serving as the foundational proof-of-concept for the subsequent Enhanced and Ultra versions.

## Technical Specifications

| Feature | Specification |
| :--- | :--- |
| **Model Architecture** | Llama 3.2 (Dense) |
| **Parameter Count** | 1.24 Billion |
| **Task Type** | Text-to-Text Generation |
| **Base Model** | unsloth/Llama-3.2-1B-Instruct-bnb-4bit |
| **Quantization** | 4-bit NormalFloat (NF4) via bitsandbytes |
| **Context Length** | 2048 Tokens |
| **Language Support** | Italian, English |

## Fine-Tuning Details (PEFT/LoRA)

*   **Method:** LoRA (Low-Rank Adaptation)
*   **Rank (r):** 16
*   **Alpha:** 16
*   **Target Modules:** `q_proj`, `k_proj`, `v_proj`, `o_proj`
*   **Dataset:** Alpaca-GPT4 (Filtered subset for logical consistency)
*   **Epochs:** 1
*   **Optimizer:** AdamW 8-bit

## Lineage & Evolution

1.  **Nexa 0.1 First (Current):** Initial alpha release focusing on base instruction following.
2.  **Nexa 0.2 Enhanced:** Optimization of response fluidity and broader vocabulary.
3.  **Nexa 0.3 Enhanced:** High-density tuning (Rank 128) for advanced reasoning.
4.  **Nexa 0.4 Enhanced:** Final optimization targeting all-linear projections for maximum parameter density.

## How to Use

```python
from unsloth import FastLanguageModel
import torch

model, tokenizer = FastLanguageModel.from_pretrained(
    model_name = "WeAreNexa/Nexa_0.1_First",
    max_seq_length = 2048,
    load_in_4bit = True,
)

# Fast inference mode