t5-small-billsum

This model is a fine-tuned version of google-t5/t5-small on Billsum dataset. It achieves the following results on the evaluation set:

• Loss: 2.3195
• Rouge1: 0.205
• Rouge2: 0.0994
• Rougel: 0.1708
• Rougelsum: 0.1704
• Gen Len: 20.0

Model description

More information needed

Intended uses & limitations

• Research: Explore summarization performance of small models on long-text datasets like BillSum.

• Applications: Legislative bill summarization (short drafts), document compression.

Not intended for production without further fine-tuning and testing.

• Short outputs: Current fine-tuning capped at 20 tokens → summaries are incomplete.

• Model size: T5-small struggles with very long inputs (BillSum often >512 tokens).

• Performance: ROUGE scores (~0.20) are well below state-of-the-art.

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 2e-05
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• optimizer: Use OptimizerNames.ADAMW_TORCH_FUSED with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: linear
• num_epochs: 10
• mixed_precision_training: Native AMP

Training results

Training Loss	Epoch	Step	Validation Loss	Rouge1	Rouge2	Rougel	Rougelsum	Gen Len
No log	1.0	62	2.7672	0.139	0.0457	0.114	0.1141	20.0
No log	2.0	124	2.5470	0.148	0.0535	0.121	0.1209	20.0
No log	3.0	186	2.4587	0.16	0.0645	0.1316	0.1311	20.0
No log	4.0	248	2.4068	0.1849	0.0833	0.1541	0.1541	20.0
No log	5.0	310	2.3747	0.1994	0.0945	0.1666	0.1663	20.0
No log	6.0	372	2.3537	0.2032	0.0976	0.1686	0.1683	20.0
No log	7.0	434	2.3364	0.203	0.0973	0.1694	0.169	20.0
No log	8.0	496	2.3259	0.2045	0.0984	0.1704	0.1701	20.0
2.7564	9.0	558	2.3217	0.2052	0.0997	0.1709	0.1704	20.0
2.7564	10.0	620	2.3195	0.205	0.0994	0.1708	0.1704	20.0

Framework versions

• Transformers 4.55.4
• Pytorch 2.8.0+cu126
• Datasets 4.0.0
• Tokenizers 0.21.4