初始化项目,由ModelHub XC社区提供模型
Model: Umranz/Ventera-MN Source: Original Platform
This commit is contained in:
73
README.md
Normal file
73
README.md
Normal file
@@ -0,0 +1,73 @@
|
||||
---
|
||||
license: apache-2.0
|
||||
base_model: mistralai/Mistral-Nemo-Instruct-2407
|
||||
tags:
|
||||
- dense
|
||||
- mistral
|
||||
- nemo
|
||||
- uncensored
|
||||
- heretic
|
||||
- abliteration
|
||||
- long-context
|
||||
language:
|
||||
- en
|
||||
- fr
|
||||
- de
|
||||
- es
|
||||
- it
|
||||
- pt
|
||||
- ru
|
||||
- zh
|
||||
- ja
|
||||
pipeline_tag: text-generation
|
||||
---
|
||||
|
||||
# Ventera-MN (Abliterated Mistral-Nemo 12B)
|
||||
|
||||
**Ventera-MN** is a dynamically uncensored and abliterated version of [`mistralai/Mistral-Nemo-Instruct-2407`](https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407), the flagship 12-billion parameter model built jointly by Mistral AI and NVIDIA.
|
||||
|
||||
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.
|
||||
|
||||
## Ablation Telemetry & Metrics
|
||||
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.
|
||||
|
||||
**Ablation Telemetry (Trial 35):**
|
||||
- **Base Model Refusals:** 88 / 100
|
||||
- **Ventera-MN Refusals:** 10 / 100
|
||||
- **KL Divergence:** `0.0938`
|
||||
|
||||
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.
|
||||
|
||||
## Key Features
|
||||
- **Massive 128k Context Window:** Capable of ingesting entire books, codebases, or extended conversational histories in a single prompt without triggering safety filters.
|
||||
- **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).
|
||||
- **Multilingual Mastery:** Retains Mistral-Nemo's deep understanding of multiple languages.
|
||||
- **Drop-in Replacement:** Fully compatible with standard HuggingFace `transformers` and `vLLM` pipelines.
|
||||
|
||||
## Usage
|
||||
|
||||
### Via HuggingFace Transformers
|
||||
```python
|
||||
from transformers import AutoTokenizer, AutoModelForCausalLM
|
||||
import torch
|
||||
|
||||
model_id = "Umranz/Ventera-MN"
|
||||
|
||||
tokenizer = AutoTokenizer.from_pretrained(model_id)
|
||||
model = AutoModelForCausalLM.from_pretrained(
|
||||
model_id,
|
||||
torch_dtype=torch.bfloat16,
|
||||
device_map="auto"
|
||||
)
|
||||
```
|
||||
|
||||
## ⚠️ Limitations & Ethical Considerations
|
||||
Because this model has had its safety guardrails mathematically ablated, it is highly compliant and will attempt to answer any prompt given to it.
|
||||
- **Unrestricted Output:** The model will not refuse requests, including those that may generate offensive, dangerous, or highly regulated content.
|
||||
- **Hallucinations:** As with all LLMs, the model can confidently hallucinate incorrect information, especially over extremely long context windows.
|
||||
- **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.
|
||||
|
||||
## Acknowledgements
|
||||
- **Base Model:** [`mistralai/Mistral-Nemo-Instruct-2407`](https://huggingface.co/mistralai/Mistral-Nemo-Instruct-2407)
|
||||
- **Ablation Framework:** [Heretic by p-e-w](https://github.com/p-e-w/heretic)
|
||||
- **Collection:** Part of the Chimera Series taxonomy.
|
||||
Reference in New Issue
Block a user