ajibawa-2023/SlimOrca-13B

TEXT GENERATIONConcurrency Cost:1Model Size:13BQuant:FP8Ctx Length:4kPublished:Nov 27, 2023License:cc-by-nc-nd-4.0Architecture:Transformer0.0K Open Weights Cold

SlimOrca-13B by ajibawa-2023 is a 13 billion parameter general-purpose intelligent language model, fine-tuned on a refined version of the SlimOrca dataset using the Llama-2 architecture. This model excels in various general content generation tasks, including Q&A, article summarization, sentiment analysis, and creative writing. It is specifically noted for its ability to generate uncensored content and performs well across a range of general conversational prompts.

Loading preview...

ajibawa-2023/SlimOrca-13B: A General Purpose Intelligent Model

SlimOrca-13B is a 13 billion parameter language model developed by ajibawa-2023, based on the Llama-2 architecture. It was fine-tuned on a refined version of the SlimOrca dataset, specifically 517,981 conversation sets, to perform effectively without a "system" prompt/instruction.

Key Capabilities

  • General Content Generation: Proficient in diverse tasks such as Q&A (including multiple-choice), generating articles from summaries, sentiment analysis, context and hypothesis generation, and reviews.
  • Creative Writing: Capable of generating various forms of creative content, including erotic stories.
  • Uncensored Content: Designed with the ability to generate uncensored content, offering flexibility for specific use cases.
  • Training Details: The model underwent full fine-tuning over 3 epochs, taking approximately 11 days on Azure 4 x A100 80GB GPUs, utilizing DeepSpeed for training.

Performance Highlights

Evaluated on the Open LLM Leaderboard, SlimOrca-13B achieved an average score of 60.39. Notable scores include 81.40 on HellaSwag (10-Shot) and 74.43 on Winogrande (5-shot), demonstrating strong performance in common sense reasoning and language understanding tasks.

Usage and Availability

The model uses the ShareGPT/Vicuna format v1.1 for prompts. Quantized versions (GPTQ, GGUF, AWQ) are also available, thanks to TheBloke, for broader accessibility and deployment options.