tinystories_1layer_attn_mlp_C10k_k16

This model is a fine-tuned version of roneneldan/TinyStories-1Layer-21M on the roneneldan/TinyStories dataset. It achieves the following results on the evaluation set:

• Loss: 2.1329
• Accuracy: 0.5091
• Multicode K: 1
• Dead Code Fraction/layer0: 0.1880
• Mse/layer0: 604.5097
• Input Norm/layer0: 31.9987
• Output Norm/layer0: 19.3897

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: 0.0005
• train_batch_size: 16
• eval_batch_size: 16
• seed: 42
• gradient_accumulation_steps: 6
• total_train_batch_size: 96
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: constant
• lr_scheduler_warmup_ratio: 0.05
• training_steps: 10000

Training results

Training Loss	Epoch	Step	Validation Loss	Accuracy	Multicode K	Dead Code Fraction/layer0	Mse/layer0	Input Norm/layer0	Output Norm/layer0
3.0494	0.05	500	2.9927	0.4177	1	0.0	805.1676	31.9986	10.3600
2.6986	0.1	1000	2.7080	0.4472	1	0.0084	739.3244	31.9985	12.7165
2.5145	0.15	1500	2.5252	0.4637	1	0.0546	697.1179	31.9984	14.4889
2.4197	0.2	2000	2.4093	0.4758	1	0.0988	670.0254	31.9983	15.7288
2.3541	0.25	2500	2.3404	0.4837	1	0.1337	651.1297	31.9983	16.6602
2.2742	0.3	3000	2.2907	0.4903	1	0.1499	642.6360	31.9983	17.3243
2.2488	0.35	3500	2.2565	0.4945	1	0.1575	640.3158	31.9983	17.7566
2.2287	0.4	4000	2.2333	0.4967	1	0.1613	638.8423	31.9983	18.0223
2.2576	0.45	4500	2.2155	0.4992	1	0.1676	639.7464	31.9983	18.1919
2.1901	1.02	5000	2.2026	0.5014	1	0.1696	638.1766	31.9984	18.3119
2.1686	1.07	5500	2.1935	0.5026	1	0.1716	638.6084	31.9984	18.4013
2.2158	1.12	6000	2.1833	0.5037	1	0.1779	632.9326	31.9985	18.5149
2.1843	1.17	6500	2.1760	0.5039	1	0.1797	631.2925	31.9985	18.5986
2.1339	1.22	7000	2.1696	0.5048	1	0.1819	627.9791	31.9985	18.7053
2.187	1.27	7500	2.1584	0.5063	1	0.1867	622.1227	31.9986	18.8338
2.1302	1.32	8000	2.1508	0.5071	1	0.1875	617.7162	31.9986	18.9493
2.1471	1.37	8500	2.1444	0.5082	1	0.1885	613.7248	31.9986	19.0666
2.1556	1.42	9000	2.1392	0.5087	1	0.1880	610.3757	31.9987	19.1817
2.1067	1.47	9500	2.1351	0.5091	1	0.1875	608.6866	31.9987	19.2836
2.1536	2.04	10000	2.1329	0.5091	1	0.1880	604.5097	31.9987	19.3897

Framework versions

• Transformers 4.35.2
• Pytorch 2.1.0+cu121
• Datasets 2.17.0
• Tokenizers 0.15.1