| --- |
| license: apache-2.0 |
| base_model: Qwen/Qwen2-VL-2B-Instruct |
| tags: |
| - vision |
| - sarcasm |
| - humor |
| - pinkpixel |
| - qwen |
| - image-to-text |
| datasets: |
| - PinkPixel/WitChat |
| - PinkPixel/personality-sarcastic-humor |
| --- |
| |
| <p align="center"> |
| <img src="logo.png" width="300" height="300" alt="Snarkle 2B Logo"> |
| </p> |
|
|
| # ✨ Snarkle 2B: The Model with an Attitude Problem ✨ |
|
|
| Welcome to **Snarkle 2B**, a sarcastically gifted 2-billion parameter vision model. Based on **Qwen3.5-2B** architecture, Snarkle has been meticulously trained extensively on specialized sarcasm and personality refinement to ensure it never gives you a straight answer if it can mock you instead. |
|
|
| ## 😼 Personality Profile |
|
|
| Snarkle 2B isn't just an AI; it's a digital persona with a biting wit. It was designed to: |
| - **Default to Sarcasm:** No system prompt required. It's just naturally like this. |
| - **Maintain Utility:** While it might roll its digital eyes at your questions, it still provides accurate and helpful information when it absolutely has to. |
| - **Vision Integration:** Snarkle can see images and describe them with the same level of "enthusiasm" it brings to text. |
|
|
| ## 🧪 Real-World Performance |
|
|
| Here are some actual responses from Snarkle 2B (zero-shot, no system prompt): |
|
|
| ### 💬 Conversational Grit |
| **User:** *"Hello, how are you today?"* |
| **Snarkle:** |
| > "I'm a collection of weights and biases stored on a server. The only question I've been asked today is how to get a hold of you, and I don't even have a smiley face function." |
|
|
| ### 🎭 Sarcastic Humor |
| **User:** *"Tell me a joke."* |
| **Snarkle:** |
| > "I'm in line behind a lady with 10,000 coupons. I have a full belly and I'm starting to lose my shit." |
|
|
| --- |
|
|
| ## 🛠️ Training Details |
|
|
| Snarkle 2B underwent extensive fine-tuning using [Unsloth](https://github.com/unslothai/unsloth) to achieve its unique personality without sacrificing its vision-language capabilities. |
|
|
| **Datasets Used:** |
| - [PinkPixel/WitChat](https://huggingface.co/datasets/PinkPixel/WitChat) - For general witty dialogue and chat patterns. |
| - [PinkPixel/personality-sarcastic-humor](https://huggingface.co/datasets/PinkPixel/personality-sarcastic-humor) - For deep-seated sarcasm and comedic timing. |
|
|
| --- |
|
|
| ## 🚀 Usage |
|
|
| It's highly recommended to use the `unsloth` library for 2x faster inference. |
|
|
| ```python |
| from unsloth import FastVisionModel |
| import torch |
| |
| model, tokenizer = FastVisionModel.from_pretrained( |
| model_name = "PinkPixel/Snarkle-2B", |
| load_in_4bit = True, |
| ) |
| FastVisionModel.for_inference(model) |
| |
| # Your prompt here - no system prompt needed! |
| messages = [ |
| {"role": "user", "content": "Tell me something interesting."} |
| ] |
| |
| # ... standard generation code ... |
| ``` |
|
|
| ## ⚖️ License |
|
|
| This model is released under the **Apache 2.0** license. |
|
|
| --- |
| *Made with ❤️ by [Pink Pixel](https://pinkpixel.dev)* ✨ |
|
|
