CodeXRyu/meeting-summarizer

language: en license: apache-2.0 tags:

• text-summarization
• meeting-summarization
• t5
• transformers
• qmsum datasets:
• qmsum metrics:
• rouge pipeline_tag: summarization

Meeting Summarizer

This model is a fine-tuned version of t5-small for meeting summarization tasks.

Model Details

• Base Model: t5-small
• Task: Abstractive Meeting Summarization
• Training Data: QMSum Dataset + Enhanced Training
• Parameters: t5-small architecture

Training Configuration

• Max Input Length: 256 tokens
• Max Output Length: 64 tokens
• Batch Size: 16
• Learning Rate: 5e-05
• Training Epochs: 1
• Training Samples: N/A

Usage

from transformers import AutoTokenizer, AutoModelForSeq2SeqLM

# Load model and tokenizer
tokenizer = AutoTokenizer.from_pretrained("CodeXRyu/meeting-summarizer")
model = AutoModelForSeq2SeqLM.from_pretrained("CodeXRyu/meeting-summarizer")

def generate_summary(meeting_text, max_length=150):
    # Prepare input
    input_text = "summarize: " + meeting_text
    inputs = tokenizer(input_text, max_length=512, truncation=True, return_tensors="pt")
    
    # Generate summary
    summary_ids = model.generate(
        inputs["input_ids"],
        max_length=max_length,
        num_beams=4,
        length_penalty=2.0,
        early_stopping=True
    )
    
    return tokenizer.decode(summary_ids[0], skip_special_tokens=True)

# Example usage
meeting_transcript = '''
John: Good morning team. Let's discuss our Q3 results.
Sarah: Our sales exceeded targets by 15%, reaching $2.1M in revenue.
Mike: The new marketing campaign was very effective.
John: Great work everyone. Let's plan for Q4.
'''

summary = generate_summary(meeting_transcript)
print(summary)

Training Data

This model was trained on the QMSum dataset, which contains real meeting transcripts from multiple domains:

• Academic meetings
• Product development meetings
• Committee meetings

Performance

The model achieves competitive ROUGE scores on meeting summarization benchmarks.

Limitations

• Optimized for English meeting transcripts
• Performance may vary on very long meetings (>512 tokens input)
• Best suited for structured meeting formats with speaker labels

Citation

If you use this model, please cite:

@misc{meeting-summarizer-codexryu,
  author = {CodeXRyu},
  title = {Meeting Summarizer},
  year = {2025},
  publisher = {Hugging Face},
  url = {https://huggingface.co/CodeXRyu/meeting-summarizer}
}