Hugging Face Overview
nbeerbower/Llama-3.1-Nemotron-lorablated-70B is a 70 billion parameter language model built upon the Llama-3.1-Nemotron architecture. Its primary distinction is its uncensored nature, achieved by merging the nvidia/Llama-3.1-Nemotron-70B-Instruct-HF base model with mlabonne/Llama-3-70B-Instruct-abliterated-LORA using a technique called task arithmetic. This process effectively ablates censorship layers from the original model.
Key Capabilities & Features
- Uncensored Responses: Designed to provide less restricted outputs compared to its base model.
- Llama-3.1-Nemotron Foundation: Benefits from the underlying capabilities of the Llama-3.1-Nemotron-70B-Instruct-HF model.
- LoRA Ablation Method: Utilizes a specific LoRA (Low-Rank Adaptation) merging technique to modify model behavior.
- High Context Length: Supports a context window of 32768 tokens.
Performance Metrics
Evaluations on the Open LLM Leaderboard show an average score of 33.69. Specific scores include:
- IFEval (0-Shot): 71.47
- BBH (3-Shot): 48.06
- MMLU-PRO (5-shot): 43.46
When to Use This Model
This model is particularly suited for use cases where an uncensored large language model is required, especially if the base Llama-3.1-Nemotron-70B-Instruct-HF model's inherent safety alignments are too restrictive for the application. Developers interested in exploring the effects of LoRA-based censorship ablation on powerful base models will find this model relevant.