qwen3-14b-comedy-reward

qwen3-14b-comedy-reward

LoRA adapter from the Stanford humor-research project (April 2026).

See https://github.com/Fchaubard/llm-humor-research for the full pipeline, training scripts, and evaluation. The headline result of that repo is that this adapter (sft_50k variant) beats Qwen2.5-14B-Instruct nucleus baseline 89 % of the time on a 15-prompt pairwise tournament judged by Gemini-Flash with symmetry probing.

Quick use

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

BASE = "Qwen/Qwen2.5-14B-Instruct"
ADAPTER = "fchaubard/qwen3-14b-comedy-reward"

tok = AutoTokenizer.from_pretrained(BASE)
bnb = BitsAndBytesConfig(load_in_4bit=True, bnb_4bit_compute_dtype=torch.bfloat16,
                          bnb_4bit_quant_type="nf4", bnb_4bit_use_double_quant=True)
model = AutoModelForCausalLM.from_pretrained(BASE, quantization_config=bnb, device_map={"": 0})
model = PeftModel.from_pretrained(model, ADAPTER, autocast_adapter_dtype=False)
model.eval()

msgs = [{"role":"system", "content":"You are a sharp comedian."},
        {"role":"user",   "content":"Tell me a one-liner about midlife crises."}]
chat = tok.apply_chat_template(msgs, tokenize=False, add_generation_prompt=True)
inp = tok([chat], return_tensors="pt").to("cuda:0")
out = model.generate(**inp, do_sample=True, top_p=0.95, temperature=0.95, max_new_tokens=64)
print(tok.decode(out[0][inp["input_ids"].shape[1]:], skip_special_tokens=True))