BART
Sequence-to-sequence model for abstractive summarization and text generation
BART (Bidirectional and Auto-Regressive Transformers) is a denoising autoencoder pre-trained for sequence-to-sequence tasks. The large-cnn variant is specifically optimized for news summarization. Can be used with base weights or a fine-tuned checkpoint.
When to use:
- Abstractive summarization of articles or documents
- Text transformation and paraphrasing
- Custom seq2seq tasks after fine-tuning
Input: Input text to summarize or transform + optional fine-tuned checkpoint Output: Generated summary or transformed text, plus metadata (lengths, tokens)
Model Settings
Model Variant (default: large-cnn, options: base / large / large-cnn) Which BART checkpoint to load.
- base: Smaller, faster, lower accuracy
- large: Full-size model, general-purpose
- large-cnn: Fine-tuned on CNN/DailyMail — best for news summarization
Inference Settings
Max Length (default: 142, range: 10–1024) Maximum token length of the generated output.
- Increase for longer documents that need fuller summaries
- Decrease for concise bullet-point style summaries
Min Length (default: 56) Minimum token length of the generated output.
- Prevents trivially short outputs
- Set lower if short summaries are acceptable
Num Beams (default: 4) Number of beams for beam search.
- 1: Greedy decoding (fastest, lowest quality)
- 4: Good balance of quality and speed
- 8+: Higher quality, slower
Length Penalty (default: 2.0) Exponential penalty applied to sequence length during beam search.
- Values > 1.0: Encourage longer outputs
- Values < 1.0: Encourage shorter outputs
Early Stopping (default: true) Stop beam search when all beams have finished generating.
- Keep true for standard use — prevents wasted computation