---
library_name: mlx-lm
base_model: mlx-community/Llama-3.2-3B-Instruct-4bit
tags:
- tech
- ai
- research papers
- twitter
- viral-content
- mlx
- lora
license: mit
language:
- en
---

# Tech Tweet Generator Llama-3 (Fine-Tuned)

This model is a fine-tuned version of **Llama-3.2-3B-Instruct** designed to convert dense scientific and technical research paper abstracts into engaging, viral Twitter threads.

It was trained using **LoRA (Low-Rank Adaptation)** on the Apple MLX framework.

## 🚀 Model Description

- **Developed by:** Meet Merchant
- **Base Model:** `mlx-community/Llama-3.2-3B-Instruct-4bit`
- **Task:** Summarization / Style Transfer (Research Paper Abstract -> Engaging Twitter Thread)
- **Language:** English
- **Framework:** MLX

## 💻 How to Use

You can run this model locally on your Mac using `mlx-lm`.

### Installation
```bash
pip install mlx-lm
```

### Python Code
```python
from mlx_lm import load, generate

# Load the model and your adapters
model, tokenizer = load(
    "mlx-community/Llama-3.2-3B-Instruct-4bit",
    adapter_path="meetmerchant/tech-tweet-generator-llama3" 
)

abstract = """
[Paste Abstract Here]
"""

prompt = f"""
<|begin_of_text|><|start_header_id|>system<|end_header_id|>
You are a viral science communicator.
<|eot_id|><|start_header_id|>user<|end_header_id|>
Title: Example Paper
Abstract:
{abstract}
<|eot_id|><|start_header_id|>assistant<|end_header_id|>
"""

response = generate(model, tokenizer, prompt=prompt, max_tokens=500)
print(response)
```

## 📊 Training Details

- **Dataset:** 50+ ArXiv papers from the field of AI.
- **Ground Truth:** Generated by GPT-4o-mini ("Teacher-Student" distillation).
- **Training Config:**
  - LoRA Rank: 16
  - Quantization: 4-bit
  - Iterations: 200

## 🏆 Evaluation

The model was evaluated using **LLM-as-a-Judge (GPT-4o)** against the base model.
- **Win Rate:** 66% (vs Base Model)
- **Strengths:** High engagement, emoji usage, accessible language.
- **Weaknesses:** Can occasionally hallucinate specific details if the abstract is too dense.

---
*Built with ❤️ on a MacBook Pro M3 using Apple MLX.*