60 lines
1.8 KiB
Markdown
60 lines
1.8 KiB
Markdown
---
|
|
language:
|
|
- en
|
|
- fr
|
|
- de
|
|
- es
|
|
- it
|
|
- pt
|
|
- ru
|
|
- zh
|
|
- ja
|
|
tags:
|
|
- abliterated
|
|
- uncensored
|
|
- ministral
|
|
- mistral
|
|
- text-generation
|
|
- conversational
|
|
base_model: mistralai/Ministral-8B-Instruct-2410
|
|
datasets:
|
|
- N/A
|
|
pipeline_tag: text-generation
|
|
model-index:
|
|
- name: ministral-8B-Instruct-2410-abliterated
|
|
results: []
|
|
---
|
|
|
|
# Ministral-8B-Instruct-2410-Abliterated
|
|
|
|
This is an **abliterated** version of Mistral AI's Ministral-8B-Instruct-2410 model. Through surgical intervention on the model's weights, this version has reduced refusal behaviors while maintaining the core capabilities of the original model.
|
|
|
|
## What is Abliteration?
|
|
|
|
Abliteration is a technique that modifies a model's internal representations to reduce built-in refusal and censorship mechanisms. This process aims to make the model more compliant with user requests while preserving its fundamental capabilities and knowledge.
|
|
|
|
## Model Details
|
|
|
|
**Base Model**: This model is derived from [mistralai/Ministral-8B-Instruct-2410](https://huggingface.co/mistralai/Ministral-8B-Instruct-2410)
|
|
|
|
**Architecture**: Same as base Ministral-8B with the following specifications:
|
|
- **8.02B parameters**
|
|
- **128k context window** with interleaved sliding-window attention
|
|
- **Vocabulary size**: 131,072 tokens using the V3-Tekken tokenizer
|
|
- **Multilingual and code capabilities**
|
|
- **Function calling support**
|
|
|
|
## Key Features
|
|
|
|
### Enhanced Responsiveness
|
|
- Reduced refusal behaviors compared to the original model
|
|
- More direct answers to complex or sensitive queries
|
|
- Maintains the base model's strong performance on general tasks
|
|
|
|
### Technical Capabilities
|
|
All original capabilities are preserved:
|
|
- Long-context understanding up to 128k tokens
|
|
- Strong multilingual performance
|
|
- Code generation and understanding
|
|
- Function calling support
|
|
- Instruction following |