dialogsum-phi3-lora

A LoRA adapter for dialogue summarization, fine-tuned on top of Phi-3-mini-4k-instruct using the DialogSum dataset.

Given a multi-turn messenger-style conversation, the model generates a concise one-paragraph summary of what was discussed and agreed upon.

Live demo on HuggingFace Spaces

---

Usage

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

base_model_id = "microsoft/Phi-3-mini-4k-instruct"
adapter_id = "rotemso23/dialogsum-phi3-lora"

tokenizer = AutoTokenizer.from_pretrained(base_model_id, trust_remote_code=True)

model = AutoModelForCausalLM.from_pretrained(
    base_model_id,
    torch_dtype=torch.float16,
    device_map="auto",
    trust_remote_code=False,
)
model = PeftModel.from_pretrained(model, adapter_id)
model.eval()

dialogue = """Amanda: I baked cookies. Do you want some?
Jerry: Sure! What kind?
Amanda: Chocolate chip. I'll bring them to the office tomorrow.
Jerry: Amazing, I can't wait. Thanks Amanda!
Amanda: No problem :)"""

messages = [{"role": "user", "content": f"Summarize the following conversation in a few sentences.\n\nConversation:\n{dialogue}"}]
prompt = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)

with torch.inference_mode():
    output = model.generate(**inputs, max_new_tokens=128, do_sample=False)

summary = tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True)
print(summary)

---

Training details

Setting	Value
Base model	microsoft/Phi-3-mini-4k-instruct
Dataset	DialogSum (4,000 train / 500 val / 819 test)
LoRA rank	16
LoRA alpha	32
Target modules	qkv_proj, o_proj
Epochs	3
Effective batch size	16
Learning rate	2e-4
Quantization	4-bit NF4 (training only)
Hardware	NVIDIA T4 (Google Colab)

---

Evaluation results

Evaluated on the full DialogSum test set (819 examples):

Model	ROUGE-1	ROUGE-2	ROUGE-L
Phi-3-mini zero-shot (baseline)	0.312	0.101	0.238
This model (fine-tuned)	0.474	0.215	0.391

---

Prompt format

<|user|>
Summarize the following conversation in a few sentences.

Conversation:
{dialogue}
<|end|>
<|assistant|>