dphn/dolphin-2.1-70b

TEXT GENERATIONConcurrency Cost:4Model Size:69BQuant:FP8Ctx Length:32kPublished:Oct 23, 2023License:llama2Architecture:Transformer0.0K Open Weights Cold

dphn/dolphin-2.1-70b is a 69 billion parameter uncensored language model developed by Eric Hartford, based on the Llama 2 architecture and fine-tuned on the StellarBright base model. It is an open-source implementation of Microsoft's Orca dataset, modified for uncensoring and enhanced with the Airoboros dataset for increased creativity. This model is designed to be highly compliant to any requests, making it suitable for use cases where an unaligned model is desired, provided an external alignment layer is implemented.

Loading preview...

Dolphin 2.1-70b: An Uncensored, Highly Compliant LLM

dphn/dolphin-2.1-70b is a 69 billion parameter language model developed by Eric Hartford, built upon the Llama 2 architecture and fine-tuned using the StellarBright base model. This model is an open-source adaptation of Microsoft's Orca dataset, with significant modifications for uncensoring, deduplication, and quality improvements. It also integrates Jon Durbin's Airoboros dataset to boost creative generation capabilities.

Key Characteristics

  • Uncensored and Highly Compliant: The model's dataset has been filtered to remove alignment and bias, making it highly compliant to a wide range of requests, including potentially unethical ones. Users are advised to implement their own alignment layer for responsible deployment.
  • Llama 2 Base: Suitable for both commercial and non-commercial applications due to its Llama 2 foundation.
  • Enhanced Creativity: The inclusion of the Airoboros dataset aims to improve the model's creative text generation.
  • Training Details: Trained for 3 epochs over 5 days on 4x A100 GPUs using qLoRA and the Axolotl framework.
  • Prompt Format: Utilizes the ChatML prompt format, with a specific instruction for the model to output "###FINISHED###" when it completes its response, as a workaround for a known bug with EOS token generation.

Important Note

This specific release (dolphin-2.1-70b) contains a known bug preventing it from generating EOS tokens. Users should consider using the updated ehartford/dolphin-2.2-70b where this issue has been resolved.

Finetunes
1 models