t5-small-lora-tweetsumm

T5-Small + LoRA — TweetSumm Dialogue Summarizer

LoRA adapter (rank 4, α=16, rsLoRA) fine-tuned on customer service dialogues from the TweetSumm dataset. Only 0.24% of parameters (147K of 60M) are trained.

• 🖥️ Live demo: lora-tweetsumm-demo
• 🧬 Training code, rank ablation and evaluation: github.com/lucianoon/lora-tweetsumm

Usage

from peft import PeftModel
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

base = AutoModelForSeq2SeqLM.from_pretrained("google-t5/t5-small")
model = PeftModel.from_pretrained(base, "lucianoon/t5-small-lora-tweetsumm")
tokenizer = AutoTokenizer.from_pretrained("google-t5/t5-small")

dialogue = "Customer: I need to cancel order #12345.\nAgent: Done! Refund in 3-5 days."
inputs = tokenizer("summarize: " + dialogue, return_tensors="pt", truncation=True, max_length=512)
outputs = model.generate(**inputs, max_new_tokens=48)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Training

Rank-4 was selected via ablation over r ∈ {4, 8, 16, 32} — see the GitHub repository for the full study (ROUGE-L 0.357 on the reference run).