athirdpath/Orca-2-13b-Alpaca-Uncensored
athirdpath/Orca-2-13b-Alpaca-Uncensored is a 13 billion parameter language model fine-tuned from Microsoft's Orca-2-13b. It was specifically adjusted on a subset of the Vezora/Mini_Orca_Uncencored_Alpaca dataset to improve instruction following and reduce content rejections. This model maintains strong reasoning capabilities for its size while offering more permissive responses compared to the original Orca-2.
Loading preview...
Model Overview
This model, athirdpath/Orca-2-13b-Alpaca-Uncensored, is a 13 billion parameter language model derived from microsoft/Orca-2-13b. It has been fine-tuned using a subset of the Vezora/Mini_Orca_Uncencored_Alpaca dataset.
Key Capabilities & Differentiators
- Instruction Following: The fine-tuning process specifically targeted the
q_projandk_projmodules with a low rank (8) to enhance the model's ability to follow instructions and adjust its prompt format. - Reduced Rejections: A primary goal of this fine-tuning was to reduce instances of content rejection, particularly for "spicy prompts," making it more permissive than the base Orca-2 model.
- Reasoning Performance: Despite the alignment adjustments, the model largely retains the solid reasoning capabilities characteristic of a 13B model.
- Performance Comparison: While slightly worse than the original Orca-2 in Ooba's chat mode, it performs comparably in Alpaca chat-instruct mode to the original in ChatLM chat-instruct mode.
Benchmarks
Evaluated on the Open LLM Leaderboard, the model achieves an average score of 61.63. Specific metric scores include:
- AI2 Reasoning Challenge (25-Shot): 61.09
- HellaSwag (10-Shot): 79.27
- MMLU (5-Shot): 60.13
- TruthfulQA (0-shot): 53.59
- Winogrande (5-shot): 77.43
- GSM8k (5-shot): 38.29
Usage Considerations
While designed to be more permissive, some "shocking prompts" may still be rejected. The creator suggests using techniques like an author's note or character card to overcome such rejections.