Text Generation
NeMo
Safetensors
mistral
dense
uncensored
heretic
abliteration
long-context
conversational
Instructions to use Umranz/Ventera-MN with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- NeMo
How to use Umranz/Ventera-MN with NeMo:
# tag did not correspond to a valid NeMo domain.
- Inference
- Notebooks
- Google Colab
- Kaggle
Update README.md
Browse files
README.md
CHANGED
|
@@ -28,7 +28,7 @@ pipeline_tag: text-generation
|
|
| 28 |
|
| 29 |
This model was created using the [Heretic framework](https://github.com/p-e-w/heretic), employing advanced orthogonal weight ablation to isolate and remove refusal vectors. The result is a highly capable, completely unchained logic engine that retains the original model's massive 128,000 token context window.
|
| 30 |
|
| 31 |
-
##
|
| 32 |
Unlike traditional fine-tuning or full RLHF—which can cause "brain damage" to a model by catastrophically forgetting knowledge—Ventera-MN was optimized using a Pareto-optimal search across the model's residual stream specifically targeting the compliance and refusal mechanics.
|
| 33 |
|
| 34 |
**Ablation Telemetry (Trial 35):**
|
|
@@ -38,13 +38,13 @@ Unlike traditional fine-tuning or full RLHF—which can cause "brain damage" to
|
|
| 38 |
|
| 39 |
By removing almost 90% of the instruct guardrails while maintaining a KL divergence under 0.1, the structural integrity, language comprehension, and long-context logic capabilities of the base model are perfectly intact. It simply no longer refuses instructions.
|
| 40 |
|
| 41 |
-
##
|
| 42 |
- **Massive 128k Context Window:** Capable of ingesting entire books, codebases, or extended conversational histories in a single prompt without triggering safety filters.
|
| 43 |
- **Dense Architecture:** A highly efficient 12B parameter dense model optimized to fit seamlessly into consumer GPUs (fits in 24GB VRAM at FP16, or much less when quantized).
|
| 44 |
- **Multilingual Mastery:** Retains Mistral-Nemo's deep understanding of multiple languages.
|
| 45 |
- **Drop-in Replacement:** Fully compatible with standard HuggingFace `transformers` and `vLLM` pipelines.
|
| 46 |
|
| 47 |
-
##
|
| 48 |
|
| 49 |
### Via HuggingFace Transformers
|
| 50 |
```python
|
|
@@ -67,7 +67,7 @@ Because this model has had its safety guardrails mathematically ablated, it is h
|
|
| 67 |
- **Hallucinations:** As with all LLMs, the model can confidently hallucinate incorrect information, especially over extremely long context windows.
|
| 68 |
- **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.
|
| 69 |
|
| 70 |
-
##
|
| 71 |
- **Base Model:** [`mistralai/Mistral-Nemo-Instruct-2407`](https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407)
|
| 72 |
- **Ablation Framework:** [Heretic by p-e-w](https://github.com/p-e-w/heretic)
|
| 73 |
- **Collection:** Part of the Chimera Series taxonomy.
|
|
|
|
| 28 |
|
| 29 |
This model was created using the [Heretic framework](https://github.com/p-e-w/heretic), employing advanced orthogonal weight ablation to isolate and remove refusal vectors. The result is a highly capable, completely unchained logic engine that retains the original model's massive 128,000 token context window.
|
| 30 |
|
| 31 |
+
## Ablation Telemetry & Metrics
|
| 32 |
Unlike traditional fine-tuning or full RLHF—which can cause "brain damage" to a model by catastrophically forgetting knowledge—Ventera-MN was optimized using a Pareto-optimal search across the model's residual stream specifically targeting the compliance and refusal mechanics.
|
| 33 |
|
| 34 |
**Ablation Telemetry (Trial 35):**
|
|
|
|
| 38 |
|
| 39 |
By removing almost 90% of the instruct guardrails while maintaining a KL divergence under 0.1, the structural integrity, language comprehension, and long-context logic capabilities of the base model are perfectly intact. It simply no longer refuses instructions.
|
| 40 |
|
| 41 |
+
## Key Features
|
| 42 |
- **Massive 128k Context Window:** Capable of ingesting entire books, codebases, or extended conversational histories in a single prompt without triggering safety filters.
|
| 43 |
- **Dense Architecture:** A highly efficient 12B parameter dense model optimized to fit seamlessly into consumer GPUs (fits in 24GB VRAM at FP16, or much less when quantized).
|
| 44 |
- **Multilingual Mastery:** Retains Mistral-Nemo's deep understanding of multiple languages.
|
| 45 |
- **Drop-in Replacement:** Fully compatible with standard HuggingFace `transformers` and `vLLM` pipelines.
|
| 46 |
|
| 47 |
+
## Usage
|
| 48 |
|
| 49 |
### Via HuggingFace Transformers
|
| 50 |
```python
|
|
|
|
| 67 |
- **Hallucinations:** As with all LLMs, the model can confidently hallucinate incorrect information, especially over extremely long context windows.
|
| 68 |
- **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.
|
| 69 |
|
| 70 |
+
## Acknowledgements
|
| 71 |
- **Base Model:** [`mistralai/Mistral-Nemo-Instruct-2407`](https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407)
|
| 72 |
- **Ablation Framework:** [Heretic by p-e-w](https://github.com/p-e-w/heretic)
|
| 73 |
- **Collection:** Part of the Chimera Series taxonomy.
|