Babsie/DavidAU_L3.3_8B_OpusR_HUC
Babsie/DavidAU_L3.3_8B_OpusR_HUC is an 8 billion parameter Llama 3.3-based instruction-tuned model developed by DavidAU, featuring a 128k context length. This model is 'Heretic'd' (de-censored) and further trained with a Claude 4.5-Opus High Reasoning dataset, creating an uncensored instruct/thinking hybrid. It is specifically designed for neuroatypical accessibility and excels at complex reasoning tasks, automatically activating a 'thinking' process for detailed responses.
Loading preview...
Model Overview
Babsie/DavidAU_L3.3_8B_OpusR_HUC is an 8 billion parameter model based on a "Llama3.3-8B" source, adjusted to a 128k context length. Developed by DavidAU, this model has undergone significant modifications to enhance its reasoning and instruction-following capabilities while being fully uncensored.
Key Capabilities
- Uncensored Output: The model has been "Heretic'd" to remove original censorship and trained to reduce refusals (14/100 refusal rate, KL divergence 0.1498).
- High Reasoning: Fine-tuned with a Claude 4.5-Opus High Reasoning dataset, it functions as an instruct/thinking hybrid.
- Automatic Thinking Activation: Certain phrases or words in prompts (e.g., "Think Deeply:") automatically activate a detailed internal "thinking" process, leading to more elaborate and structured responses.
- Extended Context: Supports a 128k context window, with a suggested minimum context window of 4k and optimal performance at 8k+.
Good For
- Complex Reasoning Tasks: Ideal for prompts requiring detailed explanations, mathematical examples, or structured creative writing where the "thinking" activation can be leveraged.
- Unrestricted Content Generation: Suitable for use cases where uncensored or explicit content generation is desired, though it may require explicit directives (e.g., using specific slang or terms) to achieve the desired graphic level.
- Neuroatypical Accessibility: Developed with a focus on accessibility for neuroatypical users.
Benchmarks
Preliminary benchmarks show scores such as ARC-Challenge: 0.480, ARC-Easy: 0.687, BoolQ: 0.831, HellaSwag: 0.705, OpenBookQA: 0.438, PIQA: 0.780, and Winogrande: 0.646.