Upload folder using huggingface_hub
Browse files- .gitattributes +1 -0
- README.md +287 -0
- qwen3-vl-2b-thinking-abliterated.gguf +3 -0
- qwen3-vl-2b-thinking-abliterated.safetensors +3 -0
.gitattributes
CHANGED
|
@@ -33,3 +33,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 33 |
*.zip filter=lfs diff=lfs merge=lfs -text
|
| 34 |
*.zst filter=lfs diff=lfs merge=lfs -text
|
| 35 |
*tfevents* filter=lfs diff=lfs merge=lfs -text
|
| 36 |
+
qwen3-vl-2b-thinking-abliterated.gguf filter=lfs diff=lfs merge=lfs -text
|
README.md
ADDED
|
@@ -0,0 +1,287 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
+
library_name: transformers
|
| 4 |
+
pipeline_tag: image-text-to-text
|
| 5 |
+
tags:
|
| 6 |
+
- qwen3
|
| 7 |
+
- vision-language
|
| 8 |
+
- multimodal
|
| 9 |
+
- abliterated
|
| 10 |
+
- thinking
|
| 11 |
+
- image-generation
|
| 12 |
+
---
|
| 13 |
+
|
| 14 |
+
<!-- README Version: v1.0 -->
|
| 15 |
+
|
| 16 |
+
# Qwen3-VL-2B-Thinking (Abliterated)
|
| 17 |
+
|
| 18 |
+
A 2-billion parameter vision-language model from the Qwen3-VL family, featuring abliterated safety filters for unrestricted generation and enhanced reasoning capabilities. This model combines visual understanding with text generation, enabling multimodal analysis and creative applications.
|
| 19 |
+
|
| 20 |
+
## Model Description
|
| 21 |
+
|
| 22 |
+
**Qwen3-VL-2B-Thinking-Abliterated** is a modified version of Qwen3-VL-2B optimized for:
|
| 23 |
+
|
| 24 |
+
- **Vision-Language Understanding**: Process images and generate contextual text responses
|
| 25 |
+
- **Multimodal Reasoning**: Analyze visual content with detailed explanations
|
| 26 |
+
- **Unrestricted Generation**: Abliterated safety layers for creative freedom
|
| 27 |
+
- **Thinking Mode**: Enhanced reasoning and step-by-step analysis capabilities
|
| 28 |
+
- **Efficient Inference**: 2B parameters for consumer hardware deployment
|
| 29 |
+
|
| 30 |
+
**Key Features**:
|
| 31 |
+
- Visual question answering (VQA)
|
| 32 |
+
- Image captioning and description
|
| 33 |
+
- Visual reasoning and analysis
|
| 34 |
+
- Multimodal conversation
|
| 35 |
+
- Creative image interpretation
|
| 36 |
+
|
| 37 |
+
## Repository Contents
|
| 38 |
+
|
| 39 |
+
```
|
| 40 |
+
qwen3-vl-2b-thinking/
|
| 41 |
+
├── qwen3-vl-2b-thinking-abliterated.safetensors # PyTorch model (4.0GB)
|
| 42 |
+
└── qwen3-vl-2b-thinking-abliterated.gguf # GGUF quantized (3.3GB)
|
| 43 |
+
```
|
| 44 |
+
|
| 45 |
+
**Total Repository Size**: ~7.3GB
|
| 46 |
+
|
| 47 |
+
### Model Formats
|
| 48 |
+
|
| 49 |
+
| File | Format | Size | Use Case |
|
| 50 |
+
|------|--------|------|----------|
|
| 51 |
+
| `qwen3-vl-2b-thinking-abliterated.safetensors` | SafeTensors | 4.0GB | Transformers, PyTorch |
|
| 52 |
+
| `qwen3-vl-2b-thinking-abliterated.gguf` | GGUF | 3.3GB | llama.cpp, Ollama |
|
| 53 |
+
|
| 54 |
+
## Hardware Requirements
|
| 55 |
+
|
| 56 |
+
### Minimum Requirements
|
| 57 |
+
- **VRAM**: 6GB (GGUF quantized inference)
|
| 58 |
+
- **RAM**: 8GB system memory
|
| 59 |
+
- **Disk Space**: 8GB available
|
| 60 |
+
- **GPU**: CUDA-compatible (RTX 2060+) or Apple Silicon
|
| 61 |
+
|
| 62 |
+
### Recommended Requirements
|
| 63 |
+
- **VRAM**: 8-12GB (SafeTensors full precision)
|
| 64 |
+
- **RAM**: 16GB system memory
|
| 65 |
+
- **Disk Space**: 10GB available
|
| 66 |
+
- **GPU**: RTX 3060 Ti+ or Apple M1 Pro+
|
| 67 |
+
|
| 68 |
+
### Performance Estimates
|
| 69 |
+
- **GGUF on 8GB VRAM**: ~15-25 tokens/sec
|
| 70 |
+
- **SafeTensors on 12GB VRAM**: ~20-35 tokens/sec
|
| 71 |
+
- **CPU inference**: ~2-5 tokens/sec (not recommended)
|
| 72 |
+
|
| 73 |
+
## Usage Examples
|
| 74 |
+
|
| 75 |
+
### Using Transformers (SafeTensors)
|
| 76 |
+
|
| 77 |
+
```python
|
| 78 |
+
from transformers import Qwen2VLForConditionalGeneration, AutoProcessor
|
| 79 |
+
from PIL import Image
|
| 80 |
+
import torch
|
| 81 |
+
|
| 82 |
+
# Load model and processor
|
| 83 |
+
model_path = "E:/huggingface/qwen3-vl-2b-thinking"
|
| 84 |
+
model = Qwen2VLForConditionalGeneration.from_pretrained(
|
| 85 |
+
model_path,
|
| 86 |
+
torch_dtype=torch.float16,
|
| 87 |
+
device_map="auto"
|
| 88 |
+
)
|
| 89 |
+
processor = AutoProcessor.from_pretrained(model_path)
|
| 90 |
+
|
| 91 |
+
# Load image
|
| 92 |
+
image = Image.open("image.jpg")
|
| 93 |
+
|
| 94 |
+
# Create conversation
|
| 95 |
+
messages = [
|
| 96 |
+
{
|
| 97 |
+
"role": "user",
|
| 98 |
+
"content": [
|
| 99 |
+
{"type": "image", "image": image},
|
| 100 |
+
{"type": "text", "text": "Describe this image in detail."}
|
| 101 |
+
]
|
| 102 |
+
}
|
| 103 |
+
]
|
| 104 |
+
|
| 105 |
+
# Process and generate
|
| 106 |
+
text = processor.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
|
| 107 |
+
inputs = processor(text=[text], images=[image], return_tensors="pt").to("cuda")
|
| 108 |
+
|
| 109 |
+
# Generate response
|
| 110 |
+
with torch.no_grad():
|
| 111 |
+
outputs = model.generate(**inputs, max_new_tokens=256)
|
| 112 |
+
response = processor.batch_decode(outputs, skip_special_tokens=True)[0]
|
| 113 |
+
|
| 114 |
+
print(response)
|
| 115 |
+
```
|
| 116 |
+
|
| 117 |
+
### Using llama.cpp (GGUF)
|
| 118 |
+
|
| 119 |
+
```bash
|
| 120 |
+
# Download llama.cpp with vision support
|
| 121 |
+
git clone https://github.com/ggerganov/llama.cpp
|
| 122 |
+
cd llama.cpp
|
| 123 |
+
make
|
| 124 |
+
|
| 125 |
+
# Run inference with image
|
| 126 |
+
./llama-cli \
|
| 127 |
+
--model "E:/huggingface/qwen3-vl-2b-thinking/qwen3-vl-2b-thinking-abliterated.gguf" \
|
| 128 |
+
--image "image.jpg" \
|
| 129 |
+
--prompt "Describe this image:" \
|
| 130 |
+
--n-gpu-layers 32 \
|
| 131 |
+
--ctx-size 4096
|
| 132 |
+
```
|
| 133 |
+
|
| 134 |
+
### Using Ollama (GGUF)
|
| 135 |
+
|
| 136 |
+
```bash
|
| 137 |
+
# Create Modelfile
|
| 138 |
+
cat > Modelfile <<EOF
|
| 139 |
+
FROM E:/huggingface/qwen3-vl-2b-thinking/qwen3-vl-2b-thinking-abliterated.gguf
|
| 140 |
+
PARAMETER temperature 0.7
|
| 141 |
+
PARAMETER top_p 0.9
|
| 142 |
+
EOF
|
| 143 |
+
|
| 144 |
+
# Create Ollama model
|
| 145 |
+
ollama create qwen3-vl-thinking -f Modelfile
|
| 146 |
+
|
| 147 |
+
# Run interactive session
|
| 148 |
+
ollama run qwen3-vl-thinking "Analyze this image: image.jpg"
|
| 149 |
+
```
|
| 150 |
+
|
| 151 |
+
### Visual Question Answering
|
| 152 |
+
|
| 153 |
+
```python
|
| 154 |
+
# Detailed analysis with thinking mode
|
| 155 |
+
messages = [
|
| 156 |
+
{
|
| 157 |
+
"role": "user",
|
| 158 |
+
"content": [
|
| 159 |
+
{"type": "image", "image": image},
|
| 160 |
+
{"type": "text", "text": "Think step-by-step and explain what's happening in this image."}
|
| 161 |
+
]
|
| 162 |
+
}
|
| 163 |
+
]
|
| 164 |
+
|
| 165 |
+
# Model will provide detailed reasoning in its response
|
| 166 |
+
```
|
| 167 |
+
|
| 168 |
+
## Model Specifications
|
| 169 |
+
|
| 170 |
+
### Architecture
|
| 171 |
+
- **Model Type**: Vision-Language Transformer
|
| 172 |
+
- **Base Architecture**: Qwen3-VL
|
| 173 |
+
- **Parameters**: 2 billion
|
| 174 |
+
- **Modifications**: Abliterated safety layers, enhanced reasoning
|
| 175 |
+
- **Vision Encoder**: ViT-based image encoder
|
| 176 |
+
- **Text Decoder**: Qwen3 transformer decoder
|
| 177 |
+
|
| 178 |
+
### Technical Details
|
| 179 |
+
- **Precision**: FP16 (SafeTensors), Quantized (GGUF)
|
| 180 |
+
- **Context Length**: 4096 tokens
|
| 181 |
+
- **Image Resolution**: 448x448 (default), up to 1024x1024
|
| 182 |
+
- **Vocabulary Size**: ~151,000 tokens
|
| 183 |
+
- **Training**: Multimodal pretraining + instruction tuning
|
| 184 |
+
|
| 185 |
+
### Supported Tasks
|
| 186 |
+
- Image captioning
|
| 187 |
+
- Visual question answering
|
| 188 |
+
- Scene understanding
|
| 189 |
+
- Object detection (descriptive)
|
| 190 |
+
- Visual reasoning
|
| 191 |
+
- Image-to-text generation
|
| 192 |
+
- Multimodal conversation
|
| 193 |
+
|
| 194 |
+
## Performance Tips
|
| 195 |
+
|
| 196 |
+
### Optimization Strategies
|
| 197 |
+
|
| 198 |
+
1. **VRAM Optimization**:
|
| 199 |
+
```python
|
| 200 |
+
# Use 8-bit quantization
|
| 201 |
+
model = Qwen2VLForConditionalGeneration.from_pretrained(
|
| 202 |
+
model_path,
|
| 203 |
+
load_in_8bit=True,
|
| 204 |
+
device_map="auto"
|
| 205 |
+
)
|
| 206 |
+
```
|
| 207 |
+
|
| 208 |
+
2. **Image Preprocessing**:
|
| 209 |
+
```python
|
| 210 |
+
# Resize large images
|
| 211 |
+
from PIL import Image
|
| 212 |
+
image = Image.open("large_image.jpg")
|
| 213 |
+
image = image.resize((448, 448))
|
| 214 |
+
```
|
| 215 |
+
|
| 216 |
+
3. **Batch Processing**:
|
| 217 |
+
```python
|
| 218 |
+
# Process multiple images efficiently
|
| 219 |
+
images = [Image.open(f"image{i}.jpg") for i in range(4)]
|
| 220 |
+
inputs = processor(images=images, text=prompts, return_tensors="pt")
|
| 221 |
+
```
|
| 222 |
+
|
| 223 |
+
4. **GGUF Performance**:
|
| 224 |
+
- Use `--n-gpu-layers 32` for GPU acceleration
|
| 225 |
+
- Adjust `--ctx-size` based on available VRAM
|
| 226 |
+
- Use `--threads` for CPU optimization
|
| 227 |
+
|
| 228 |
+
### Generation Parameters
|
| 229 |
+
|
| 230 |
+
```python
|
| 231 |
+
generation_config = {
|
| 232 |
+
"max_new_tokens": 256,
|
| 233 |
+
"temperature": 0.7,
|
| 234 |
+
"top_p": 0.9,
|
| 235 |
+
"do_sample": True,
|
| 236 |
+
"repetition_penalty": 1.1
|
| 237 |
+
}
|
| 238 |
+
|
| 239 |
+
outputs = model.generate(**inputs, **generation_config)
|
| 240 |
+
```
|
| 241 |
+
|
| 242 |
+
## Abliteration Notice
|
| 243 |
+
|
| 244 |
+
This model has been **abliterated** (safety filters removed) for:
|
| 245 |
+
- Unrestricted creative content generation
|
| 246 |
+
- Research and experimentation
|
| 247 |
+
- Artistic applications without content restrictions
|
| 248 |
+
|
| 249 |
+
**Important**: Users are responsible for ethical use and compliance with local laws. This model may generate unrestricted content.
|
| 250 |
+
|
| 251 |
+
## License
|
| 252 |
+
|
| 253 |
+
Licensed under **Apache 2.0**. Free for commercial and research use with attribution.
|
| 254 |
+
|
| 255 |
+
Key provisions:
|
| 256 |
+
- ✅ Commercial use permitted
|
| 257 |
+
- ✅ Modification and distribution allowed
|
| 258 |
+
- ✅ Private use permitted
|
| 259 |
+
- ⚠️ Provide attribution and license notice
|
| 260 |
+
- ⚠️ State changes if modified
|
| 261 |
+
|
| 262 |
+
Full license: [Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0)
|
| 263 |
+
|
| 264 |
+
## Citation
|
| 265 |
+
|
| 266 |
+
```bibtex
|
| 267 |
+
@misc{qwen3vl2b-thinking-abliterated,
|
| 268 |
+
title={Qwen3-VL-2B-Thinking-Abliterated},
|
| 269 |
+
author={Qwen Team and Community Contributors},
|
| 270 |
+
year={2025},
|
| 271 |
+
howpublished={\url{https://huggingface.co/Qwen}},
|
| 272 |
+
note={Abliterated vision-language model with enhanced reasoning}
|
| 273 |
+
}
|
| 274 |
+
```
|
| 275 |
+
|
| 276 |
+
## Official Resources
|
| 277 |
+
|
| 278 |
+
- **Qwen Official**: https://github.com/QwenLM/Qwen
|
| 279 |
+
- **Transformers Docs**: https://huggingface.co/docs/transformers
|
| 280 |
+
- **llama.cpp**: https://github.com/ggerganov/llama.cpp
|
| 281 |
+
- **Model Family**: https://huggingface.co/Qwen
|
| 282 |
+
|
| 283 |
+
## Acknowledgments
|
| 284 |
+
|
| 285 |
+
- **Qwen Team**: Original Qwen3-VL architecture and pretraining
|
| 286 |
+
- **Community**: Abliteration techniques and reasoning enhancements
|
| 287 |
+
- **Hugging Face**: Model hosting and transformers library
|
qwen3-vl-2b-thinking-abliterated.gguf
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:7530922c695bb15091a39e690f1379461338b630cb271816aee051bed8bf13f3
|
| 3 |
+
size 3447350144
|
qwen3-vl-2b-thinking-abliterated.safetensors
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:293ac2e5469f4350e3f8b5484fbdb95d49e2be534cdc2408f976ac88332d29d6
|
| 3 |
+
size 4255140312
|