Overview

This is an experimental 7 billion parameter language model developed by diffnamehard, fine-tuned from the Mistral-CatMacaroni-slerp-7B base model. Its primary characteristic is its training on the toxic-dpo-v0.1-NoWarning-alpaca dataset, which aims to produce uncensored outputs.

Key Capabilities

• General Language Understanding: The model exhibits capabilities across a range of common language understanding tasks.
• Uncensored Responses: Fine-tuning on a specific dataset suggests an emphasis on generating responses without typical content filters.
• Benchmark Performance: Achieves competitive scores on several academic benchmarks, including:
  • ARC (25-shot): 64.25
  • HellaSwag (10-shot): 84.09
  • MMLU (5-shot): 62.66
  • TruthfulQA (0-shot): 56.87
  • Winogrande (5-shot): 79.72
  • GSM8K (5-shot): 56.1

Good For

• Experimental Use Cases: Ideal for researchers and developers exploring the behavior of models fine-tuned on specific, less-filtered datasets.
• Comparative Analysis: Useful for comparing performance and response characteristics against other models, particularly those with standard safety alignments.
• Applications Requiring Directness: Potentially suitable for applications where direct, unfiltered language generation is a requirement, provided ethical considerations are thoroughly addressed.