Model Card for phi-2-dialogsum

This model is designed for dialogue summarization. It takes multi-turn conversations as input and produces concise summaries.

Model Details

Model Description

This is the model card for phi-2-dialogsum, a dialogue summarization model built on top of 🤗 Transformers. It leverages phi-2 backbone model, fine-tuned for summarizing dialogues.

Developed by: Aygün Varol & Malik Sami
Model type: Generative Language Model
Language(s) (NLP): English
License: MIT
Finetuned from model: Phi-2

Model Sources

Repository (GitHub): AygunVarol/phi-2-dialogsum

Uses

Direct Use

This model can be used directly for dialogue summarization tasks. For example, given a multi-turn conversation, the model will produce a succinct summary capturing the key information and context.

How to Get Started with the Model

Below is a quick code snippet to load and run inference with this model:

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

model_name = "YourHuggingFaceUsername/phi-2-dialogsum"  # replace with the correct HF model ID
tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForSeq2SeqLM.from_pretrained(model_name)

input_text = """Speaker1: Hi, how are you doing today?
Speaker2: I'm good, thanks! Just finished my coffee.
Speaker1: That's nice. Did you sleep well last night?
Speaker2: Actually, I slept quite late watching a new show on Netflix."""
inputs = tokenizer([input_text], max_length=512, truncation=True, return_tensors="pt")

summary_ids = model.generate(**inputs, max_length=60, num_beams=4, early_stopping=True)
summary = tokenizer.decode(summary_ids[0], skip_special_tokens=True)

print("Summary:", summary)

Training Details

Training dataset Dialogsum

Evaluation

ORIGINAL MODEL: {'rouge1': 0.2990526195120211, 'rouge2': 0.10874019046839419, 'rougeL': 0.21186900909813286, 'rougeLsum': 0.22342464591439556}

PEFT MODEL: {'rouge1': 0.3132817683433486, 'rouge2': 0.1070363134080079, 'rougeL': 0.23226760188839027, 'rougeLsum': 0.25947902747914586}

Absolute percentage improvement of PEFT MODEL over ORIGINAL MODEL

rouge1: 1.42%

rouge2: -0.17%

rougeL: 2.04%

rougeLsum: 3.61%