pszemraj/mega-ar-small-4096-wikitext-103-raw-v1

mega-ar-small-4096: MWE

mega-ar on wikitext-103-raw-v1 (document level)

This model is a fine-tuned version of pszemraj/random-mega-ar-small-4096 on the EleutherAI/wikitext_document_level dataset (wikitext-103-raw-v1). This model has ~ 65M params.

It achieves the following results on the evaluation set:

• Loss: 4.0338
• Accuracy: 0.3243

Training procedure

This was tuned with bf16, while the authors recommend tuning with fp32. Will try fp32 later.

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 4
• eval_batch_size: 1
• seed: 80085
• gradient_accumulation_steps: 32
• total_train_batch_size: 128
• optimizer: Adam with betas=(0.9,0.98) and epsilon=1e-07
• lr_scheduler_type: linear
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy
7.3662	0.11	100	7.2782	0.0935
6.3064	0.22	200	6.2066	0.1634
5.8203	0.33	300	5.7299	0.1931
5.55	0.44	400	5.4173	0.2117
5.3194	0.55	500	5.1937	0.2278
5.1678	0.66	600	5.0206	0.2406
5.0375	0.77	700	4.8891	0.2508
4.9194	0.88	800	4.7592	0.2605
4.8272	0.99	900	4.6653	0.2681
4.7571	1.1	1000	4.5817	0.2754
4.6345	1.21	1100	4.5066	0.2820
4.6218	1.32	1200	4.4472	0.2867
4.5585	1.43	1300	4.3827	0.2923
4.5047	1.54	1400	4.3328	0.2963
4.4726	1.65	1500	4.2860	0.3002
4.4094	1.76	1600	4.2452	0.3038
4.3705	1.87	1700	4.2168	0.3062
4.3739	1.98	1800	4.1852	0.3095
4.2836	2.09	1900	4.1599	0.3112
4.302	2.2	2000	4.1307	0.3149
4.2847	2.31	2100	4.1113	0.3165
4.2348	2.42	2200	4.0925	0.3184
4.2837	2.53	2300	4.0743	0.3207
4.2058	2.64	2400	4.0612	0.3217
4.22	2.75	2500	4.0494	0.3224
4.1827	2.86	2600	4.0397	0.3237
4.1967	2.97	2700	4.0338	0.3243

Framework versions

• Transformers 4.32.1
• Pytorch 2.1.0.dev20230727+cu118
• Datasets 2.13.1
• Tokenizers 0.13.3