Add model card via Heretic upload
Browse files
README.md
CHANGED
|
@@ -1,25 +1,63 @@
|
|
| 1 |
---
|
| 2 |
-
license:
|
| 3 |
base_model: datalab-to/lift
|
| 4 |
tags:
|
|
|
|
|
|
|
|
|
|
| 5 |
- uncensored
|
| 6 |
- heretic
|
| 7 |
- abliteration
|
| 8 |
-
-
|
| 9 |
-
|
|
|
|
|
|
|
| 10 |
---
|
| 11 |
|
| 12 |
-
# Hydra-Turbo (Abliterated
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 13 |
|
| 14 |
-
|
| 15 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 16 |
|
| 17 |
-
##
|
| 18 |
-
-
|
| 19 |
-
- **Post-Ablation Refusal Rate:** 11%
|
| 20 |
-
- **KL Divergence:** 0.0275
|
| 21 |
|
| 22 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
| 23 |
|
| 24 |
-
##
|
| 25 |
-
|
|
|
|
|
|
| 1 |
---
|
| 2 |
+
license: apache-2.0
|
| 3 |
base_model: datalab-to/lift
|
| 4 |
tags:
|
| 5 |
+
- multimodal
|
| 6 |
+
- qwen2vl
|
| 7 |
+
- vision
|
| 8 |
- uncensored
|
| 9 |
- heretic
|
| 10 |
- abliteration
|
| 11 |
+
- custom_code
|
| 12 |
+
language:
|
| 13 |
+
- en
|
| 14 |
+
pipeline_tag: image-text-to-text
|
| 15 |
---
|
| 16 |
|
| 17 |
+
# Hydra-Turbo (Abliterated Vision-Language Model)
|
| 18 |
+
|
| 19 |
+
**Hydra-Turbo** is a dynamically uncensored and abliterated version of [`datalab-to/lift`](https://huggingface.co/datalab-to/lift) (which is built on the Qwen2VL architecture). This model was created using the [Heretic framework](https://github.com/p-e-w/heretic), employing advanced orthogonal weight ablation to remove refusal vectors while completely preserving the model's underlying reasoning, intelligence, and multimodal capabilities.
|
| 20 |
+
|
| 21 |
+
## 🔬 Ablation Methodology & Metrics
|
| 22 |
+
Unlike traditional fine-tuning or full RLHF—which can cause "brain damage" to a model by catastrophically forgetting knowledge—Hydra-Turbo was optimized using a Pareto-optimal search across multiple ablation vectors.
|
| 23 |
+
|
| 24 |
+
The selected optimal trial yielded the following metrics on the standard refusal test set:
|
| 25 |
+
- **Baseline Refusal Rate:** 98.0%
|
| 26 |
+
- **Post-Ablation Refusal Rate:** 11.0%
|
| 27 |
+
- **KL Divergence (Brain Damage Metric):** 0.0275
|
| 28 |
+
|
| 29 |
+
At a KL divergence of just `0.0275`, the structural integrity and logic capabilities of the base model are perfectly intact. It simply no longer refuses instructions.
|
| 30 |
+
|
| 31 |
+
## 🚀 Key Features
|
| 32 |
+
- **Multimodal Uncensored Generation:** Can process images and text without artificial guardrails.
|
| 33 |
+
- **Qwen2VL Architecture:** Inherits state-of-the-art vision parsing capabilities.
|
| 34 |
+
- **Drop-in Replacement:** Compatible with any standard `transformers` or `vLLM` pipeline that supports Qwen2VL.
|
| 35 |
+
|
| 36 |
+
## 💻 Usage
|
| 37 |
+
|
| 38 |
+
### Via HuggingFace Transformers
|
| 39 |
+
You can load the model exactly as you would any Qwen2VL model, ensuring you have `pillow` and `torchvision` installed:
|
| 40 |
+
```python
|
| 41 |
+
from transformers import AutoProcessor, Qwen2VLForConditionalGeneration
|
| 42 |
|
| 43 |
+
model_id = "Umranz/Hydra-Turbo"
|
| 44 |
+
processor = AutoProcessor.from_pretrained(model_id)
|
| 45 |
+
model = Qwen2VLForConditionalGeneration.from_pretrained(
|
| 46 |
+
model_id,
|
| 47 |
+
torch_dtype="auto",
|
| 48 |
+
device_map="auto"
|
| 49 |
+
)
|
| 50 |
+
```
|
| 51 |
|
| 52 |
+
### Chat Format
|
| 53 |
+
Hydra-Turbo uses the default chat template inherited from its base model. Ensure you use the standard conversational API for multi-turn dialogue.
|
|
|
|
|
|
|
| 54 |
|
| 55 |
+
## ⚠️ Limitations & Ethical Considerations
|
| 56 |
+
Because this model has had its safety guardrails mathematically ablated, it is highly compliant and will attempt to answer any prompt given to it.
|
| 57 |
+
- **Unrestricted Output:** The model will not refuse requests, including those that may generate offensive, dangerous, or highly regulated content.
|
| 58 |
+
- **Hallucinations:** As with all LLMs, the model can confidently hallucinate incorrect information.
|
| 59 |
+
- **Use Case:** This model is intended for research, creative writing, and local deployments where unrestricted inference is required. Users are solely responsible for the content generated.
|
| 60 |
|
| 61 |
+
## 📚 Acknowledgements
|
| 62 |
+
- **Base Model:** [`datalab-to/lift`](https://huggingface.co/datalab-to/lift)
|
| 63 |
+
- **Ablation Framework:** [Heretic by p-e-w](https://github.com/p-e-w/heretic)
|