johngiorgi's picture
Make doc sep token visible
482bf76
metadata
language: en
tags:
  - generated_from_trainer
datasets:
  - allenai/mslr2022
model-index:
  - name: baseline
    results: []

Overview

This model is a fine-tuned version of allenai/led-base-16384 on the Cochrane dataset. The model received as input the titles and abstracts of up to 25 included studies for each example, concatenated by the "</s>" token. Global attention is applied to the special start token "<s>" and each of the document separator tokens "</s>". The model slightly outperforms the reported results in the original paper: MS2: Multi-Document Summarization of Medical Studies. See the Cochrane leaderboard for results on the blind test set.

It achieves the following results on the validation set:

  • Loss: 4.0216
  • Rouge1 Fmeasure Mean: 26.3026
  • Rouge2 Fmeasure Mean: 6.0324
  • Rougel Fmeasure Mean: 18.1513
  • Rougelsum Fmeasure Mean: 22.5031
  • Bertscore Hashcode: microsoft/deberta-xlarge-mnli_L40_no-idf_version=0.3.11(hug_trans=4.22.0.dev0)-rescaled_fast-tokenizer
  • Bertscore F1 Mean: 20.5937
  • Seed: 42
  • Model Name Or Path: allenai/led-base-16384
  • Doc Sep Token: "</s>"

Model description

More information needed

Intended uses & limitations

More information needed

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

  • learning_rate: 3e-05
  • train_batch_size: 4
  • eval_batch_size: 4
  • seed: 42
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 16
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_ratio: 0.1
  • num_epochs: 20
  • mixed_precision_training: Native AMP
  • label_smoothing_factor: 0.1

Framework versions

  • Transformers 4.22.0.dev0
  • Pytorch 1.12.0
  • Datasets 2.4.0
  • Tokenizers 0.12.1