Centrum / README.md
ratishsp
updates
e9e32bd
metadata
license: apache-2.0
tags:
  - generated_from_trainer
datasets:
  - NewSHead
model-index:
  - name: Centrum
    results: []

Centrum

Centrum is a pretrained model for multi-document summarization, trained with centroid-based pretraining objective on the NewSHead dataset. It is initialized from allenai/led-base-16384. The details of the approach are mentioned in the preprint Multi-Document Summarization with Centroid-Based Pretraining (Ratish Puduppully and Mark Steedman). It achieves the following results on the evaluation set:

  • Loss: 3.5568

Model description

The script for training and inference of Centrum is available on https://github.com/ratishsp/centrum

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: 1
  • eval_batch_size: 4
  • seed: 42
  • distributed_type: multi-GPU
  • num_devices: 4
  • gradient_accumulation_steps: 4
  • total_train_batch_size: 16
  • total_eval_batch_size: 16
  • optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
  • lr_scheduler_type: linear
  • lr_scheduler_warmup_steps: 10000
  • training_steps: 100000
  • mixed_precision_training: Native AMP
  • label_smoothing_factor: 0.1

Training results

Training Loss Epoch Step Validation Loss
4.1628 0.05 500 4.0732
4.0278 0.09 1000 3.9800
4.0008 0.14 1500 3.9283
3.9564 0.19 2000 3.8941
3.9193 0.23 2500 3.8780
3.9185 0.28 3000 3.8501
3.8881 0.32 3500 3.8334
3.8869 0.37 4000 3.8211
3.876 0.42 4500 3.8057
3.8552 0.46 5000 3.7954
3.8198 0.51 5500 3.7861
3.8016 0.56 6000 3.7750
3.8033 0.6 6500 3.7651
3.7927 0.65 7000 3.7528
3.7978 0.7 7500 3.7429
3.7727 0.74 8000 3.7367
3.7634 0.79 8500 3.7275
3.7395 0.83 9000 3.7158
3.7432 0.88 9500 3.7066
3.7623 0.93 10000 3.7039
3.7182 0.97 10500 3.6904
3.7146 1.02 11000 3.6881
3.681 1.07 11500 3.6797
3.6745 1.11 12000 3.6750
3.6794 1.16 12500 3.6748
3.6802 1.21 13000 3.6696
3.665 1.25 13500 3.6609
3.6516 1.3 14000 3.6633
3.6577 1.34 14500 3.6573
3.6409 1.39 15000 3.6519
3.6691 1.44 15500 3.6490
3.6521 1.48 16000 3.6475
3.6435 1.53 16500 3.6465
3.6466 1.58 17000 3.6392
3.644 1.62 17500 3.6419
3.6347 1.67 18000 3.6347
3.6205 1.71 18500 3.6328
3.6451 1.76 19000 3.6310
3.6327 1.81 19500 3.6284
3.6166 1.85 20000 3.6267
3.622 1.9 20500 3.6212
3.6164 1.95 21000 3.6199
3.6178 1.99 21500 3.6201
3.5892 2.04 22000 3.6201
3.5855 2.09 22500 3.6221
3.5658 2.13 23000 3.6193
3.5916 2.18 23500 3.6144
3.5767 2.22 24000 3.6101
3.5809 2.27 24500 3.6115
3.5561 2.32 25000 3.6110
3.5831 2.36 25500 3.6080
3.5551 2.41 26000 3.6121
3.5588 2.46 26500 3.6072
3.5645 2.5 27000 3.6056
3.5804 2.55 27500 3.6038
3.5712 2.6 28000 3.6052
3.5494 2.64 28500 3.6014
3.582 2.69 29000 3.5995
3.5487 2.73 29500 3.6051
3.5709 2.78 30000 3.5954
3.5546 2.83 30500 3.5941
3.5525 2.87 31000 3.5952
3.5603 2.92 31500 3.5972
3.5572 2.97 32000 3.5947
3.5106 3.01 32500 3.5952
3.5142 3.06 33000 3.5937
3.506 3.11 33500 3.5965
3.515 3.15 34000 3.5932
3.5247 3.2 34500 3.5951
3.5384 3.24 35000 3.5917
3.5165 3.29 35500 3.5887
3.5187 3.34 36000 3.5866
3.5097 3.38 36500 3.5895
3.5136 3.43 37000 3.5878
3.5095 3.48 37500 3.5839
3.5226 3.52 38000 3.5859
3.5277 3.57 38500 3.5827
3.4959 3.62 39000 3.5846
3.5003 3.66 39500 3.5823
3.5095 3.71 40000 3.5820
3.4814 3.75 40500 3.5854
3.5173 3.8 41000 3.5796
3.4968 3.85 41500 3.5810
3.5183 3.89 42000 3.5783
3.512 3.94 42500 3.5784
3.5069 3.99 43000 3.5775
3.5014 4.03 43500 3.5819
3.4787 4.08 44000 3.5836
3.4625 4.12 44500 3.5788
3.4902 4.17 45000 3.5784
3.4927 4.22 45500 3.5773
3.4813 4.26 46000 3.5769
3.4637 4.31 46500 3.5761
3.4731 4.36 47000 3.5771
3.4856 4.4 47500 3.5786
3.4579 4.45 48000 3.5790
3.5032 4.5 48500 3.5738
3.4826 4.54 49000 3.5749
3.4709 4.59 49500 3.5746
3.4916 4.63 50000 3.5745
3.4715 4.68 50500 3.5706
3.4926 4.73 51000 3.5729
3.4974 4.77 51500 3.5725
3.4796 4.82 52000 3.5683
3.4817 4.87 52500 3.5707
3.4683 4.91 53000 3.5721
3.4986 4.96 53500 3.5689
3.4763 5.01 54000 3.5716
3.4668 5.05 54500 3.5700
3.4274 5.1 55000 3.5724
3.4499 5.14 55500 3.5717
3.4507 5.19 56000 3.5706
3.4343 5.24 56500 3.5697
3.4151 5.28 57000 3.5710
3.4469 5.33 57500 3.5712
3.458 5.38 58000 3.5692
3.4559 5.42 58500 3.5680
3.4354 5.47 59000 3.5683
3.4479 5.52 59500 3.5703
3.4627 5.56 60000 3.5678
3.4478 5.61 60500 3.5659
3.4645 5.65 61000 3.5675
3.4658 5.7 61500 3.5666
3.4657 5.75 62000 3.5658
3.4618 5.79 62500 3.5653
3.4541 5.84 63000 3.5653
3.4552 5.89 63500 3.5648
3.4679 5.93 64000 3.5648
3.4423 5.98 64500 3.5652
3.3893 6.03 65000 3.5646
3.4239 6.07 65500 3.5668
3.4329 6.12 66000 3.5639
3.4151 6.16 66500 3.5649
3.4181 6.21 67000 3.5682
3.4314 6.26 67500 3.5669
3.4245 6.3 68000 3.5629
3.421 6.35 68500 3.5663
3.4329 6.4 69000 3.5660
3.4122 6.44 69500 3.5651
3.4362 6.49 70000 3.5628
3.4497 6.54 70500 3.5648
3.431 6.58 71000 3.5626
3.432 6.63 71500 3.5648
3.4208 6.67 72000 3.5635
3.4526 6.72 72500 3.5645
3.4139 6.77 73000 3.5621
3.4212 6.81 73500 3.5629
3.4352 6.86 74000 3.5597
3.4242 6.91 74500 3.5597
3.429 6.95 75000 3.5619
3.4133 7.0 75500 3.5592
3.4086 7.04 76000 3.5621
3.4056 7.09 76500 3.5604
3.4158 7.14 77000 3.5629
3.4153 7.18 77500 3.5609
3.4155 7.23 78000 3.5621
3.4117 7.28 78500 3.5626
3.407 7.32 79000 3.5638
3.3977 7.37 79500 3.5604
3.4134 7.42 80000 3.5611
3.4403 7.46 80500 3.5630
3.4002 7.51 81000 3.5601
3.4147 7.55 81500 3.5577
3.4068 7.6 82000 3.5588
3.4165 7.65 82500 3.5613
3.409 7.69 83000 3.5596
3.4213 7.74 83500 3.5583
3.403 7.79 84000 3.5601
3.3819 7.83 84500 3.5580
3.4182 7.88 85000 3.5570
3.4099 7.93 85500 3.5570
3.3845 7.97 86000 3.5582
3.411 8.02 86500 3.5610
3.3952 8.06 87000 3.5588
3.4211 8.11 87500 3.5588
3.4171 8.16 88000 3.5570
3.3825 8.2 88500 3.5607
3.3807 8.25 89000 3.5579
3.3842 8.3 89500 3.5583
3.3809 8.34 90000 3.5596
3.4033 8.39 90500 3.5590
3.4156 8.44 91000 3.5577
3.3927 8.48 91500 3.5585
3.4041 8.53 92000 3.5596
3.4006 8.57 92500 3.5600
3.4007 8.62 93000 3.5578
3.4047 8.67 93500 3.5572
3.3904 8.71 94000 3.5571
3.3888 8.76 94500 3.5581
3.3876 8.81 95000 3.5572
3.3872 8.85 95500 3.5575
3.3753 8.9 96000 3.5577
3.3961 8.95 96500 3.5568
3.4131 8.99 97000 3.5579
3.3647 9.04 97500 3.5573
3.3792 9.08 98000 3.5576
3.3755 9.13 98500 3.5575
3.3981 9.18 99000 3.5573
3.3914 9.22 99500 3.5573
3.4136 9.27 100000 3.5575

Framework versions

  • Transformers 4.20.0.dev0
  • Pytorch 1.11.0
  • Datasets 2.2.2
  • Tokenizers 0.12.1