File size: 2,843 Bytes
fce0955 c4f4223 fce0955 c4f4223 fce0955 57d276a fce0955 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 | ---
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)* ✨
|