2stacks/gemma3-12b-it-comedy-v3
The 2stacks/gemma3-12b-it-comedy-v3 is a 12 billion parameter instruction-tuned Gemma 3 model, fine-tuned by 2stacks using QLoRA on a specialized comedy-style instruction dataset. This model is specifically designed to generate stand-up-style jokes, emulating the comedic voices of Mitch Hedberg, Dave Attell, and Anthony Jeselnik, along with 30 other comedians. It excels at producing comedic responses to user prompts, trading general helpfulness for a distinct humorous voice.
Loading preview...
Overview
2stacks/gemma3-12b-it-comedy-v3 is a 12 billion parameter Gemma 3 instruction-tuned model, fine-tuned by 2stacks using QLoRA. It was trained on the 2stacks/comedy-style-instruct dataset, which includes 316 examples of verbatim stand-up material and original jokes in various comedic styles. The fine-tuning aimed to develop comedic logic rather than just cadence, scaling up from a previous 4B base model.
Key Capabilities
- Comedic Voice Generation: Specialized in producing stand-up-style jokes.
- Comedian Emulation: Trained to mimic the distinct styles of Mitch Hedberg, Dave Attell, and Anthony Jeselnik, with broader coverage of 30 additional comedians.
- Instruction-Following: Responds to user prompts by generating humorous content.
Training Details
The model was fine-tuned using QLoRA with specific hyperparameters (r=32, alpha=64, dropout 0) over 4 epochs, preserving base model capabilities. It utilizes a sequence length of 1024 tokens and was trained on unsloth/gemma-3-12b-it as its base.
Caveats and Limitations
- Joke-by-Default: This model prioritizes comedic output over general helpfulness; it is not intended for general-purpose tasks.
- Dark Humor Tendency: Due to the influence of comedians like Jeselnik and Attell, the model may produce edgier or darker humor even from innocent prompts.
- Non-Commercial License: The model is licensed under CC-BY-NC-4.0, restricting its use to research, education, and personal projects only, in accordance with its underlying dataset.