AakashakaAkku/roastbot-qwen2.5-1.5b-lora

RoastBot — a Qwen2.5-1.5B roast-comedian fine-tune

RoastBot is a playful, PG-13 stand-up "roast comedian" chatbot. It was created by fine-tuning [Qwen2.5-1.5B-Instruct](https://huggingface.co/Qwen/Qwen2.5-1.5B-Instruct) (by Alibaba Cloud) with a LoRA adapter on a small, hand-crafted roast dataset.

This repository contains the LoRA adapter (my contribution). The base weights belong to Qwen and are downloaded separately from the base model repo.

🙏 Attribution

• —Base model: Qwen2.5-1.5B-Instruct — © Alibaba Cloud, licensed under Apache-2.0.
• —Fine-tune: LoRA adapter (rank 16) trained by the author to give the model a roast-comedian persona.
• —This work is released under Apache-2.0, inheriting and preserving the base model's license and attribution.

What was done

• —Authored a clean, PG-13 roast dataset (job / habit / day roasts + safety refusals that decline to roast protected traits).
• —Trained a LoRA adapter with PEFT + TRL for a consistent roast persona.
• —Added inference guardrails: English-only decoding and safety filters.

Usage

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel

base_id = "Qwen/Qwen2.5-1.5B-Instruct"
tok = AutoTokenizer.from_pretrained(base_id)
base = AutoModelForCausalLM.from_pretrained(base_id, torch_dtype=torch.float16)
model = PeftModel.from_pretrained(base, "AakashakaAkku/roastbot-qwen2.5-1.5b-lora")

msgs = [
    {"role": "system", "content": "You are RoastBot, a sharp stand-up roast comedian. Roast the user, PG-13, never cruel about things they can't change."},
    {"role": "user", "content": "I'm a product manager. Roast me."},
]
text = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
inp = tok(text, return_tensors="pt").to(model.device)
out = model.generate(**inp, max_new_tokens=110, do_sample=False, repetition_penalty=1.3)
print(tok.decode(out[0][inp.input_ids.shape[1]:], skip_special_tokens=True))

Intended use & limitations

For entertainment. It's a small (1.5B) model, so humor is hit-or-miss and it can be factually wrong. Roaster mode refuses to mock protected traits (race, gender, body, disability, etc.).

License

Apache-2.0 — please retain attribution to the Qwen base model.