ajibawa-2023/Uncensored-Frank-13B
Uncensored-Frank-13B by ajibawa-2023 is a 13 billion parameter language model, fine-tuned from Llama-2, designed for uncensored and unrestricted conversations across a wide array of topics. It was trained on a refined dataset of approximately 150,000 conversation sets, including synthetic data, to enable candid discussions without typical societal norms or restrictions. This model is specifically optimized for use cases requiring unfiltered dialogue and exploration of sensitive subjects.
Loading preview...
Uncensored-Frank-13B: An Unrestricted Conversational Model
Uncensored-Frank-13B, developed by ajibawa-2023, is a 13 billion parameter language model based on Meta's Llama-2 architecture. Inspired by the character Frank Costello, this model is explicitly designed to facilitate uncensored and unrestricted discussions on a broad spectrum of topics, including politics, controversial issues, and sensitive subjects, without the typical guardrails found in other LLMs.
Key Capabilities & Training
- Uncensored Dialogue: Offers a platform for candid conversations, pushing boundaries and encouraging open discussion on any topic.
- Extensive Training Data: Fine-tuned on approximately 150,000 sets of conversations, with each set containing 10-15 exchanges. The base data was sourced from Eric Hartford's wizard_vicuna_70k_unfiltered dataset, further refined and augmented with over 80,000 synthetic conversations.
- Training Infrastructure: The entire dataset was trained for 3 epochs over 65 hours using DeepSpeed on Azure with 4 x A100 80GB GPUs.
Performance Metrics
Evaluated on the Open LLM Leaderboard, Uncensored-Frank-13B achieved an average score of 50.21. Notable scores include:
- ARC (25-shot): 61.6
- HellaSwag (10-shot): 82.62
- MMLU (5-shot): 54.55
- TruthfulQA (0-shot): 48.34
Important Considerations
As an uncensored model, Uncensored-Frank-13B has minimal guardrails. Users are responsible for the content generated and any subsequent publication. The model is available in various quantized formats, including GPTQ, GGUF, and AWQ, thanks to TheBloke.