Multi-Domain-Expert-Learning/expert-min-pile-instruct-v1.1

layer_4,5,6,7,8

This model is a fine-tuned version of P1ayer-1/pythia-deduped-1b-chat-base on the pile-instruct/ dataset. It achieves the following results on the evaluation set:

• Loss: 4.9648
• Accuracy: 0.3842

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: 0.0001
• train_batch_size: 12
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 96
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• training_steps: 6000

Training results

Training Loss	Epoch	Step	Accuracy	Validation Loss
7.4574	0.1	200	0.1688	7.4961
7.0445	0.2	400	0.1997	7.0547
6.7483	0.3	600	0.2190	6.7930
6.4568	0.4	800	0.2376	6.5703
6.2865	0.5	1000	0.2552	6.375
6.1028	0.6	1200	0.2793	6.1484
5.8888	0.7	1400	0.2982	5.9570
5.7362	0.8	1600	0.3121	5.8008
5.6507	0.9	1800	0.3238	5.6797
5.565	1.0	2000	0.3318	5.5781
5.4688	1.1	2200	0.3392	5.4961
5.4044	1.2	2400	0.3456	5.4219
5.3323	1.3	2600	0.3516	5.3594
5.2598	1.4	2800	0.3562	5.3047
5.2159	1.5	3000	0.3596	5.2578
5.1992	1.6	3200	0.3638	5.2148
5.1429	1.69	3400	0.3672	5.1797
5.095	1.79	3600	0.3696	5.1445
5.0646	1.89	3800	0.3715	5.1172
5.059	1.99	4000	0.3742	5.0859
5.0152	2.09	4200	0.3756	5.0664
5.0251	2.19	4400	0.3769	5.0469
5.022	2.29	4600	0.3790	5.0273
4.9939	2.39	4800	0.3798	5.0156
4.924	2.49	5000	0.3811	5.0
4.9335	2.59	5200	0.3821	4.9883
4.9231	2.69	5400	0.3829	4.9805
4.8886	2.79	5600	0.3835	4.9727
4.9419	2.89	5800	0.3843	4.9648
4.9227	2.99	6000	0.3842	4.9648

Framework versions

• Transformers 4.28.1
• Pytorch 2.0.0+cu117
• Datasets 2.11.0
• Tokenizers 0.13.3

Wandb Report

https://wandb.ai/ontocord/pythia-1b-deduped-layer-test-min-pile-instruct/runs/zad9qli2