CoolFace
Modelpublic

2stacks/gemma3-12b-it-comedy-v3

sourceHugging Facecc-by-nc-4.0updated 4mo agoView on Hugging Face
0likes5downloads
Model Card

gemma3-12b-it-comedy-v3

QLoRA fine-tune of unsloth/gemma-3-12b-it on `2stacks/comedy-style-instruct` (316 examples: 120 verbatim H/A/J + 96 30-comedian variety + 100 in-the-style-of originals).

v3 scales up from the v2 4B base to test whether comedic logic (not just cadence) emerges with more model capacity. LoRA hyperparameters were backed off (r=32/α=64/4ep vs v2's r=64/α=128/6ep) to preserve base model capabilities — v2 over-fit and damaged arithmetic.

This model is trained to respond to user prompts with stand-up-style jokes, with a particular emphasis on the voices of Mitch Hedberg, Dave Attell, and Anthony Jeselnik. Style coverage extends to 30 additional comedians via the variety set.

Training

Baseunsloth/gemma-3-12b-it
MethodQLoRA r=32, alpha=64, dropout 0
Targetsq,k,v,o,gate,up,down
Schedule4 epochs, lr 0.0001, cosine, warmup 5
Batch2×4 effective 8
Seq len1024
Hardware1×H100 on Modal
Final loss4.7038

W&B: gemma3-comedy-qlora / run gemma3-12b-it-r32-a64-4ep-316ex-v3.

Files

  • —LoRA adapter (peft format)
  • —*.safetensors — merged 16-bit
  • —*.Q4_K_M.gguf — llama.cpp / Ollama format

Use

python
from transformers import AutoModelForCausalLM, AutoTokenizer
m = AutoModelForCausalLM.from_pretrained("2stacks/gemma3-12b-it-comedy-v3")
t = AutoTokenizer.from_pretrained("2stacks/gemma3-12b-it-comedy-v3")

Or in Ollama via the GGUF artifact.

Caveats

  • —Joke-by-default. This model trades general helpfulness for comedic voice. Use it for jokes; use the base model for tasks.
  • —Dark humor over-represented. Jeselnik / Attell / Stanhope material pushes the distribution toward edgier output. Expect the model to take dark turns even on innocent prompts.
  • —Non-commercial license. Per the underlying dataset, this model is CC-BY-NC-4.0 — research, education, and personal use only.

Attribution

The training data is sourced from publicly-available stand-up material released by 33 working comedians. Per-special and per-comedian attribution tables are maintained on the dataset card.

If you enjoy the voices this model imitates, please support those comedians by buying or streaming their specials directly.