A QLoRA adapter that bends Qwen2.5-7B-Instruct toward the voice of a late-night talk-show host, modeled on Trevor Noah's comedic register — global/outsider lens, calm-to-incredulous builds, character act-outs, warmth over cruelty, and a sharper human turn at the end.

This is an adapter only — it loads on top of the base model at runtime.

Why fine-tuning (not RAG/prompting)?

We're not teaching the model facts — we're teaching it a voice. Persona / style / tone is a legitimate sweet spot for fine-tuning. Facts belong in retrieval; a consistent, generalizing voice is what fine-tuning is for.

Training

Method: QLoRA (4-bit base + LoRA), rank 16, alpha 16, lr 2e-4, 3 epochs
Data: 118 chat-format training rows, single constant system prompt, loss on assistant turns only
Base: Qwen/Qwen2.5-7B-Instruct

Code, dataset, and the base-vs-tuned eval harness: https://github.com/thearpitgupta/comedian-finetune

Usage

python

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = "Qwen/Qwen2.5-7B-Instruct"
tok = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(base, device_map="auto")
model = PeftModel.from_pretrained(model, "inKMKHn/comedian-lora-v2")

system = ("You are a late-night talk show host in the style of Trevor Noah. "
          "You riff on real news headlines with a global, outsider's perspective...")
msgs = [{"role": "system", "content": system},
        {"role": "user", "content": "<a news headline>"}]
inputs = tok.apply_chat_template(msgs, add_generation_prompt=True, return_tensors="pt").to(model.device)
print(tok.decode(model.generate(inputs, max_new_tokens=300, temperature=0.9)[0]))

Caveats

Models a comedic register, not the real person. Don't clone Trevor Noah's actual voice/face — right-of-publicity and platform-ToS issues.
Trained on ~118 rows; the gap over the strong-prompted base widens with more data.
Bits are original and paraphrased from types of news events, never verbatim monologue text.