🚀 T5-Small Dialogue Summarizer (Fine-Tuned on SAMSum)
A sequence-to-sequence transformer model fine-tuned to convert conversational logs, transcripts, team syncs, and multi-turn messenger discussions into crisp, third-person narrative summaries.
📌 Model Details
- Model Architecture: Sequence-to-Sequence Transformer (
T5-small) - Base Model:
google-t5/t5-small(~60.5 Million parameters) - Training Method: Full Parameter Supervised Fine-Tuning (SFT)
- Task Conditioning Prefix:
summarize: - Target Domain: Messenger chat threads, customer support tickets, and team standup dialogues
- Language: English (
en)
🎯 Intended Uses & Limitations
💡 Intended Use
- Condensing conversational dialogues and meeting transcripts into structured key takeaways.
- Translating first-person statements (
"I fixed it") into clear third-person reporting ("Alex fixed the issue"). - Extracting actionable outcomes, timelines, bug fixes, and meeting consensus from chat threads.
⚠️ Limitations
- Speaker Formatting: Best results occur when turns follow standard conversational labeling (e.g.,
Speaker: message). - Sequence Truncation: Input sequences are bounded at 512 tokens; dialogues exceeding this cutoff require chunking to avoid dropping later context.
- Abstractive Hallucination: May occasionally misattribute specific timestamps or numerical figures if conversational threads are ambiguous.
💻 How to Use
🐍 Standard Transformers Pipeline
import torch
from transformers import AutoModelForSeq2SeqLM, AutoTokenizer
# Set your Hugging Face repository identifier
repo_id = "YOUR_HF_USERNAME/t5-small-samsum"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForSeq2SeqLM.from_pretrained(repo_id)
# Sample dialogue input
dialogue = """
Hassan: Did you get the chance to verify the memory leak fix?
Sara: Yes, I ran stress tests against the auth service. Pod memory stayed below 45%.
Hassan: Awesome. Can you deploy it to staging and ping the QA channel?
Sara: Done. Build 142 is live on staging now.
"""
# Format input with task prefix
input_text = f"summarize: {dialogue.strip()}"
inputs = tokenizer(input_text, return_tensors="pt", max_length=512, truncation=True)
# Generate summary using beam search
with torch.no_grad():
output_tokens = model.generate(
inputs["input_ids"],
max_new_tokens=60,
min_length=10,
num_beams=4,
early_stopping=True,
)
summary = tokenizer.decode(output_tokens[0], skip_special_tokens=True)
print("Summary:\n", summary)