buachaill78/dolphin-2_6-phi-2
buachaill78/dolphin-2_6-phi-2 is a 3 billion parameter language model developed by Eric Hartford and Fernando Fernandes, based on Microsoft's Phi-2 architecture. This model is specifically fine-tuned to be uncensored and highly compliant to user requests, making it suitable for applications requiring unfiltered responses. It achieves an average score of 61.7 on the Open LLM Leaderboard benchmarks, including 58.07 on GSM8K and 55.38 on MMLU.
Loading preview...
Model Overview
buachaill78/dolphin-2_6-phi-2 is a 3 billion parameter language model developed by Eric Hartford and Fernando Fernandes, built upon Microsoft's Phi-2 architecture. This iteration, version 2.6, incorporates fixes to training configuration and reintroduces Samantha-based empathy data, while replacing Synthia and Pure-Dove with Capybara in its dataset. The model was trained for 3 epochs over 2 days on 4x A100 GPUs using qLoRA and Axolotl, with sponsorship from Convai.
Key Characteristics
- Uncensored and Compliant: The model's dataset has been filtered to remove alignment and bias, resulting in an uncensored model that is highly compliant to any requests, including potentially unethical ones. Users are advised to implement their own alignment layers.
- Performance Benchmarks: On the Open LLM Leaderboard (benchmarked with lm-evaluation-harness v.0.4.1), it achieved an average score of 61.7. Notable scores include 59.81 on ARC (25-shot), 74.65 on HellaSwag (10-shot), 55.38 on MMLU (5-shot), and 58.07 on GSM8K (5-shot).
- Prompt Format: Utilizes the ChatML prompt format, expecting
system,user, andassistantroles.
Use Cases
This model is particularly suited for applications where an uncensored and highly compliant AI is desired, allowing for broad and unfiltered response generation. Developers should be aware of its uncensored nature and implement appropriate safeguards for public-facing services.