Centrum-Large / README.md
Ratish Puduppully
initial commit
edd0ba2
metadata
license: apache-2.0
tags:
  - generated_from_trainer
datasets:
  - ratishsp/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-large-16384. The details of the approach are mentioned in the ACL 2023 Multi-Document Summarization with Centroid-Based Pretraining (Ratish Puduppully, Parag Jain, Nancy F. Chen and Mark Steedman). It achieves the following results on the evaluation set:

  • Loss: 3.3292

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: 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
3.7884 0.05 500 3.7054
3.6593 0.09 1000 3.6245
3.6425 0.14 1500 3.5841
3.6008 0.19 2000 3.5561
3.5645 0.23 2500 3.5372
3.568 0.28 3000 3.5187
3.5408 0.32 3500 3.5045
3.5447 0.37 4000 3.4951
3.5324 0.42 4500 3.4845
3.5192 0.46 5000 3.4739
3.4841 0.51 5500 3.4684
3.4703 0.56 6000 3.4604
3.4759 0.6 6500 3.4534
3.4647 0.65 7000 3.4476
3.4726 0.7 7500 3.4399
3.4522 0.74 8000 3.4332
3.4454 0.79 8500 3.4277
3.4281 0.83 9000 3.4229
3.4341 0.88 9500 3.4173
3.4563 0.93 10000 3.4161
3.4188 0.97 10500 3.4094
3.3967 1.02 11000 3.4123
3.3647 1.07 11500 3.4061
3.3604 1.11 12000 3.4011
3.3662 1.16 12500 3.4011
3.3698 1.21 13000 3.3918
3.3558 1.25 13500 3.3910
3.3421 1.3 14000 3.3891
3.3468 1.34 14500 3.3894
3.3333 1.39 15000 3.3817
3.3545 1.44 15500 3.3803
3.3411 1.48 16000 3.3784
3.3338 1.53 16500 3.3782
3.3354 1.58 17000 3.3749
3.3341 1.62 17500 3.3714
3.3302 1.67 18000 3.3677
3.3179 1.71 18500 3.3659
3.3381 1.76 19000 3.3645
3.3223 1.81 19500 3.3619
3.3079 1.85 20000 3.3593
3.3156 1.9 20500 3.3576
3.3056 1.95 21000 3.3582
3.3117 1.99 21500 3.3552
3.2522 2.04 22000 3.3550
3.2522 2.09 22500 3.3586
3.2386 2.13 23000 3.3548
3.2574 2.18 23500 3.3544
3.239 2.22 24000 3.3566
3.2468 2.27 24500 3.3528
3.2264 2.32 25000 3.3511
3.2501 2.36 25500 3.3482
3.2204 2.41 26000 3.3506
3.2302 2.46 26500 3.3526
3.2353 2.5 27000 3.3492
3.2494 2.55 27500 3.3452
3.2423 2.6 28000 3.3455
3.2233 2.64 28500 3.3447
3.2498 2.69 29000 3.3420
3.2175 2.73 29500 3.3457
3.2398 2.78 30000 3.3402
3.2242 2.83 30500 3.3421
3.2185 2.87 31000 3.3457
3.2274 2.92 31500 3.3419
3.2251 2.97 32000 3.3449
3.1507 3.01 32500 3.3518
3.165 3.06 33000 3.3462
3.1512 3.11 33500 3.3434
3.1598 3.15 34000 3.3433
3.1728 3.2 34500 3.3445
3.1838 3.24 35000 3.3456
3.1649 3.29 35500 3.3442
3.1684 3.34 36000 3.3404
3.1587 3.38 36500 3.3406
3.1586 3.43 37000 3.3442
3.1545 3.48 37500 3.3381
3.1674 3.52 38000 3.3436
3.1717 3.57 38500 3.3373
3.147 3.62 39000 3.3408
3.1462 3.66 39500 3.3374
3.156 3.71 40000 3.3382
3.1354 3.75 40500 3.3366
3.1613 3.8 41000 3.3317
3.143 3.85 41500 3.3347
3.1667 3.89 42000 3.3353
3.1597 3.94 42500 3.3341
3.1566 3.99 43000 3.3357
3.124 4.03 43500 3.3410
3.1035 4.08 44000 3.3434
3.0881 4.12 44500 3.3411
3.1131 4.17 45000 3.3379
3.1191 4.22 45500 3.3468
3.1119 4.26 46000 3.3356
3.0957 4.31 46500 3.3417
3.1024 4.36 47000 3.3380
3.1141 4.4 47500 3.3472
3.0851 4.45 48000 3.3513
3.1252 4.5 48500 3.3351
3.1125 4.54 49000 3.3423
3.1019 4.59 49500 3.3396
3.1185 4.63 50000 3.3349
3.1042 4.68 50500 3.3350
3.1153 4.73 51000 3.3345
3.1289 4.77 51500 3.3356
3.1075 4.82 52000 3.3335
3.1151 4.87 52500 3.3385
3.094 4.91 53000 3.3292
3.1272 4.96 53500 3.3349
3.0847 5.01 54000 3.3407
3.0662 5.05 54500 3.3378
3.0345 5.1 55000 3.3481
3.0611 5.14 55500 3.3410
3.0566 5.19 56000 3.3424
3.0413 5.24 56500 3.3466
3.0291 5.28 57000 3.3453
3.0569 5.33 57500 3.3491
3.0645 5.38 58000 3.3378
3.0646 5.42 58500 3.3434
3.045 5.47 59000 3.3418
3.0551 5.52 59500 3.3426
3.0706 5.56 60000 3.3378
3.0556 5.61 60500 3.3407
3.0743 5.65 61000 3.3520
3.0764 5.7 61500 3.3320
3.0723 5.75 62000 3.3352
3.0716 5.79 62500 3.3327
3.0618 5.84 63000 3.3447
3.0662 5.89 63500 3.3312
3.0758 5.93 64000 3.3323
3.0501 5.98 64500 3.3400
2.978 6.03 65000 3.3473
3.0131 6.07 65500 3.3440
3.0212 6.12 66000 3.3401
3.0095 6.16 66500 3.3361
3.0118 6.21 67000 3.3352
3.0249 6.26 67500 3.3398
3.0107 6.3 68000 3.3444
3.0175 6.35 68500 3.3490
3.0241 6.4 69000 3.3402
3.0094 6.44 69500 3.3437
3.0286 6.49 70000 3.3355
3.0391 6.54 70500 3.3385
3.0243 6.58 71000 3.3395
3.0232 6.63 71500 3.3370
3.0168 6.67 72000 3.3458
3.0432 6.72 72500 3.3400
3.0121 6.77 73000 3.3420
3.0137 6.81 73500 3.3436
3.0333 6.86 74000 3.3362
3.0194 6.91 74500 3.3355
3.0198 6.95 75000 3.3434
3.0105 7.0 75500 3.3346
2.9833 7.04 76000 3.3492
2.9876 7.09 76500 3.3351
2.9918 7.14 77000 3.3466
2.9983 7.18 77500 3.3422
2.9893 7.23 78000 3.3364
2.9946 7.28 78500 3.3365
2.9851 7.32 79000 3.3402
2.9797 7.37 79500 3.3450
2.9888 7.42 80000 3.3423
3.0182 7.46 80500 3.3429
2.983 7.51 81000 3.3345
2.9959 7.55 81500 3.3397
2.9935 7.6 82000 3.3389
3.0008 7.65 82500 3.3442
2.9898 7.69 83000 3.3418
2.9989 7.74 83500 3.3387
2.985 7.79 84000 3.3482
2.963 7.83 84500 3.3369
3.0009 7.88 85000 3.3355
2.9925 7.93 85500 3.3434
2.9616 7.97 86000 3.3346
2.9769 8.02 86500 3.3430
2.9663 8.06 87000 3.3407
2.9872 8.11 87500 3.3448
2.9892 8.16 88000 3.3354
2.9526 8.2 88500 3.3445
2.9426 8.25 89000 3.3405
2.9528 8.3 89500 3.3466
2.9541 8.34 90000 3.3434
2.9643 8.39 90500 3.3475
2.9893 8.44 91000 3.3434
2.9655 8.48 91500 3.3433
2.9735 8.53 92000 3.3416
2.9722 8.57 92500 3.3443
2.9639 8.62 93000 3.3410
2.972 8.67 93500 3.3407
2.9586 8.71 94000 3.3393
2.9591 8.76 94500 3.3412
2.9523 8.81 95000 3.3411
2.9572 8.85 95500 3.3393
2.9435 8.9 96000 3.3414
2.9667 8.95 96500 3.3392
2.9824 8.99 97000 3.3428
2.9265 9.04 97500 3.3417
2.9409 9.08 98000 3.3435
2.9387 9.13 98500 3.3425
2.9635 9.18 99000 3.3420
2.9527 9.22 99500 3.3421
2.9755 9.27 100000 3.3430

Framework versions

  • Transformers 4.23.0.dev0
  • Pytorch 1.12.1
  • Datasets 2.6.1
  • Tokenizers 0.13.1