bart-large-dialogue-summarization

This model is a fine-tuned version of facebook/bart-large for English dialogue-to-summary generation.
The goal of this model is to generate concise English summaries from multi-speaker English dialogues.

Model description

This model takes an English dialogue as input and generates an English abstractive summary.
It was fine-tuned for a dialogue summarization task using paired examples with the following structure:

• Input: dialogue
• Target: summary

The model is intended for research and educational use, especially for experiments on dialogue summarization with BART.

Intended use

This model is suitable for:

• English dialogue summarization
• Research experiments on seq2seq fine-tuning
• Educational demonstrations of BART fine-tuning for summarization

This model is not intended for high-stakes or production use without further evaluation.

Training data

The model was fine-tuned on JSON files containing dialogue-summary pairs:

• train.json
• val.json
• test.json

Only the following fields were used during training:

• dialogue
• summary

Other fields such as multilingual summaries were excluded during preprocessing.

After cleaning, the final dataset sizes were:

• Train: 14,730
• Validation: 818
• Test: 819

Preprocessing

The following preprocessing steps were applied:

• Kept only dialogue and summary
• Removed unused fields such as summary_zh and summary_de
• Removed placeholder tokens such as <file_gif>
• Normalized whitespace
• Removed empty rows
• Removed duplicate rows

Training setup

The model was fine-tuned from facebook/bart-large using the Hugging Face transformers library.

Main configuration

• Base model: facebook/bart-large
• Task: English dialogue summarization
• Max source length: 768
• Max target length: 64
• Learning rate: 2e-5
• Optimizer: Adafactor
• Beam size for generation: 5
• Epochs: 5
• Per-device train batch size: 4
• Per-device eval batch size: 4
• Gradient accumulation steps: 6
• Effective train batch size: 24
• Best model selection metric: rougeLsum

Note: Some settings may be adjusted across experiments. This model card reflects the main fine-tuning configuration used for the uploaded checkpoint.

Evaluation

The model was evaluated using ROUGE on the validation and test sets.

Validation Results

Metric	Score
ROUGE-1	54.4213
ROUGE-2	30.4178
ROUGE-L	45.5119
ROUGE-Lsum	50.3035
Loss	1.3063

Test Results

Metric	Score
ROUGE-1	53.0642
ROUGE-2	28.3399
ROUGE-L	44.3812
ROUGE-Lsum	48.8576
Loss	1.3370

Example

Input

Hannah: Hey, do you have Betty's number?
Amanda: Lemme check
Hannah: <file_gif>
Amanda: Sorry, can't find it.
Amanda: Ask Larry
Amanda: He called her last time we were at the park together
Hannah: I don't know him well
Hannah: <file_gif>
Amanda: Don't be shy, he's very nice
Hannah: If you say so..
Hannah: I'd rather you texted him
Amanda: Just text him 🙂
Hannah: Urgh.. Alright
Hannah: Bye
Amanda: Bye bye

Generated Summary

Hannah is looking for Betty's number. Amanda suggests Hannah to ask Larry, who called Betty last time they were at the park.

Limitations

• The model was fine-tuned only for English dialogue summarization.
• It may hallucinate details or omit important context in long conversations.
• Performance may degrade on domains very different from the training data.
• This model should not be used in safety-critical or high-stakes settings without additional evaluation.

Citation

If you use this model, please also cite the original BART paper:

@article{lewis2019bart,
  title={BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension},
  author={Lewis, Mike and Liu, Yinhan and Goyal, Naman and Ghazvininejad, Marjan and Mohamed, Abdelrahman and Levy, Omer and Stoyanov, Veselin and Zettlemoyer, Luke},
  journal={arXiv preprint arXiv:1910.13461},
  year={2019}
}