GPT-2_para3M

This model is a pretrained version of gpt2 on an Tinystory dataset. It achieves the following results on the evaluation set:

Loss: 2.3207

Model description

More information needed

Intended uses & limitations

The limitation of this model are mainly 2 aspects.

The number of parameter of the model is only around 3.6 million which is not large. As a result the model cannot generate text in all perspectives.
The dataset is only composed of stories, this greatly hinder the performance of the model. Only stories can be generated.

Training and evaluation data

More information needed

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

learning_rate: 0.0005
train_batch_size: 32
eval_batch_size: 32
seed: 42
gradient_accumulation_steps: 8
total_train_batch_size: 256
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_steps: 100
num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
9.6976	0.01	100	7.7754
6.488	0.02	200	5.7795
5.3705	0.03	300	4.8609
4.5632	0.04	400	4.2544
4.141	0.05	500	3.9425
3.902	0.06	600	3.7189
3.7074	0.07	700	3.5514
3.5716	0.08	800	3.4291
3.4695	0.08	900	3.3253
3.3847	0.09	1000	3.2311
3.2974	0.1	1100	3.1595
3.2318	0.11	1200	3.0909
3.1698	0.12	1300	3.0329
3.1258	0.13	1400	2.9879
3.0802	0.14	1500	2.9396
3.046	0.15	1600	2.9017
3.0047	0.16	1700	2.8652
2.9701	0.17	1800	2.8320
2.9425	0.18	1900	2.8048
2.9141	0.19	2000	2.7757
2.8896	0.2	2100	2.7515
2.8667	0.21	2200	2.7263
2.8443	0.22	2300	2.7066
2.8288	0.23	2400	2.6815
2.8044	0.24	2500	2.6620
2.7886	0.25	2600	2.6471
2.7732	0.25	2700	2.6283
2.7576	0.26	2800	2.6101
2.7479	0.27	2900	2.5978
2.7256	0.28	3000	2.5819
2.7179	0.29	3100	2.5688
2.707	0.3	3200	2.5595
2.6921	0.31	3300	2.5471
2.6809	0.32	3400	2.5329
2.6779	0.33	3500	2.5232
2.663	0.34	3600	2.5154
2.6554	0.35	3700	2.5030
2.6437	0.36	3800	2.4967
2.6346	0.37	3900	2.4859
2.6293	0.38	4000	2.4768
2.6221	0.39	4100	2.4709
2.6178	0.4	4200	2.4623
2.6076	0.41	4300	2.4586
2.6025	0.41	4400	2.4492
2.5907	0.42	4500	2.4409
2.5896	0.43	4600	2.4369
2.5816	0.44	4700	2.4316
2.5783	0.45	4800	2.4256
2.577	0.46	4900	2.4204
2.5685	0.47	5000	2.4150
2.567	0.48	5100	2.4093
2.5564	0.49	5200	2.4059
2.5556	0.5	5300	2.4012
2.5496	0.51	5400	2.3997
2.545	0.52	5500	2.3956
2.5473	0.53	5600	2.3905
2.5389	0.54	5700	2.3856
2.5373	0.55	5800	2.3818
2.5318	0.56	5900	2.3787
2.5313	0.57	6000	2.3751
2.5285	0.58	6100	2.3722
2.5318	0.58	6200	2.3687
2.5229	0.59	6300	2.3666
2.5194	0.6	6400	2.3632
2.5174	0.61	6500	2.3598
2.5169	0.62	6600	2.3567
2.511	0.63	6700	2.3552
2.5093	0.64	6800	2.3546
2.5114	0.65	6900	2.3528
2.5064	0.66	7000	2.3492
2.507	0.67	7100	2.3483
2.502	0.68	7200	2.3445
2.4964	0.69	7300	2.3448
2.4999	0.7	7400	2.3423
2.4961	0.71	7500	2.3407
2.489	0.72	7600	2.3386
2.4926	0.73	7700	2.3384
2.4919	0.74	7800	2.3365
2.491	0.74	7900	2.3349
2.4893	0.75	8000	2.3333
2.4909	0.76	8100	2.3318
2.4862	0.77	8200	2.3305
2.4884	0.78	8300	2.3299
2.49	0.79	8400	2.3280
2.4788	0.8	8500	2.3286
2.4865	0.81	8600	2.3272
2.4823	0.82	8700	2.3263
2.4844	0.83	8800	2.3255
2.4826	0.84	8900	2.3251
2.4844	0.85	9000	2.3243
2.4798	0.86	9100	2.3231
2.4864	0.87	9200	2.3231
2.4755	0.88	9300	2.3228
2.4735	0.89	9400	2.3228
2.4786	0.9	9500	2.3224
2.4791	0.91	9600	2.3222
2.4809	0.91	9700	2.3214
2.4778	0.92	9800	2.3213
2.4777	0.93	9900	2.3211
2.4798	0.94	10000	2.3209
2.4768	0.95	10100	2.3212
2.4808	0.96	10200	2.3209
2.4762	0.97	10300	2.3208
2.4778	0.98	10400	2.3208
2.4816	0.99	10500	2.3207
2.4728	1.0	10600	2.3207

Framework versions

Transformers 4.32.0
Pytorch 2.0.1+cu117
Datasets 2.14.4
Tokenizers 0.13.2