led-base-big-patent

This model is a fine-tuned version of allenai/led-base-16384 on the big_patent dataset. It achieves the following results on the evaluation set:

• Loss: 0.2412
• Rouge1: 0.2972
• Rouge2: 0.1193
• Rougel: 0.2431
• Rougelsum: 0.243
• Gen Len: 46.326

Model description

This model is a long-form text generation model based on the Longformer Encoder-Decoder (LED) architecture. The LED architecture extends the Transformer model to handle long documents by incorporating sparse attention mechanisms. This makes it suitable for tasks such as summarization of lengthy patent documents, where traditional models might struggle with context length limitations. The model has been fine-tuned on the BigPatent dataset, a large collection of patent documents, to enhance its performance in generating concise and informative summaries.

Intended uses & limitations

Intended uses

• Patent summarization: Generate concise summaries of patent documents.
• Long document summarization: Useful for summarizing other types of long-form documents beyond patents.

Limitations

• Context length: Although LED handles long documents better than standard Transformers, extremely lengthy documents might still present challenges.

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 4.892476e-05
• train_batch_size: 2
• eval_batch_size: 8
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 3
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum	Gen Len
0.2195	1.0	9595	0.2350	0.2858	0.1157	0.2372	0.237	50.5544
0.1754	2.0	19191	0.2363	0.2895	0.1172	0.2392	0.2389	49.5847
0.1428	3.0	28785	0.2412	0.2972	0.1193	0.2431	0.243	46.326

Framework versions

• Transformers 4.38.2
• Pytorch 2.2.1+cu121
• Datasets 2.18.0
• Tokenizers 0.15.2