2stacks/gemma3-4b-it-comedy-v2
The 2stacks/gemma3-4b-it-comedy-v2 is a 4.3 billion parameter Gemma 3-4b-it QLoRA fine-tune by 2stacks, specifically trained to generate stand-up comedy in the style of various comedians. It excels at producing jokes with a particular emphasis on the voices of Mitch Hedberg, Dave Attell, and Anthony Jeselnik. This model is optimized for comedic text generation, trading general helpfulness for a distinct humorous voice.
Loading preview...
Overview
2stacks/gemma3-4b-it-comedy-v2 is a 4.3 billion parameter model, fine-tuned using QLoRA on the unsloth/gemma-3-4b-it base model. Its primary function is to generate stand-up comedy-style jokes, drawing from a diverse dataset of 316 examples, including verbatim material from specific comedians and a broader variety set.
Key Capabilities
- Comedic Voice Generation: Specializes in imitating the stand-up styles of Mitch Hedberg, Dave Attell, and Anthony Jeselnik, with coverage extending to 30 additional comedians.
- Joke-by-Default: Designed to prioritize comedic output over general conversational helpfulness.
- Dark Humor Tendency: Due to its training data, the model may produce edgier or darker humor, even from innocent prompts.
Training Details
The model was trained using QLoRA with specific parameters (r=64, alpha=128, dropout 0) targeting q, k, v, o, gate, up, and down layers. It underwent 6 epochs with a learning rate of 0.0001 and a cosine schedule, achieving a final loss of 3.8498. The training utilized a sequence length of 1024 and was performed on 1xH100 hardware.
Usage and Limitations
This model is intended for joke generation and is not suitable for general-purpose tasks. It operates under a CC-BY-NC-4.0 non-commercial license, restricting its use to research, education, and personal projects. Users should be aware of its predisposition towards dark humor.