---
license: cc-by-nc-4.0
base_model: unsloth/gemma-3-4b-it
datasets:
  - 2stacks/comedy-style-instruct
tags:
  - qlora
  - unsloth
  - gemma-3
  - humor
  - comedy
  - style-transfer
language:
  - en
---

# gemma3-4b-it-comedy-v2

QLoRA fine-tune of `unsloth/gemma-3-4b-it` on
[`2stacks/comedy-style-instruct`](https://huggingface.co/datasets/2stacks/comedy-style-instruct)
(316 examples: 120 verbatim H/A/J + 96 30-comedian variety + 100
in-the-style-of originals).

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

| | |
|---|---|
| Base | `unsloth/gemma-3-4b-it` |
| Method | QLoRA r=64, alpha=128, dropout 0 |
| Targets | q,k,v,o,gate,up,down |
| Schedule | 6 epochs, lr 0.0001, cosine, warmup 5 |
| Batch | 2×4 effective 8 |
| Seq len | 1024 |
| Hardware | 1×H100 on Modal |
| Final loss | 3.8498 |

W&B: `gemma3-comedy-qlora` / run `gemma3-4b-it-r64-a128-6ep-316ex-v2`.

## 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/gemma3-4b-it-comedy-v2")
t = AutoTokenizer.from_pretrained("2stacks/gemma3-4b-it-comedy-v2")
```

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](https://huggingface.co/datasets/2stacks/comedy-style-instruct).

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