medusa-ELYZA-japanese-Llama-2-7b-instruct

This model is a fine-tuned version of elyza/ELYZA-japanese-Llama-2-7b-instruct on the None dataset. It achieves the following results on the evaluation set:

• Loss: 2.3564

Model description

This is a Medusa-2 created using Medusa.

Intended uses & limitations

• 【Orion-14B Series】 Models Community License Agreement

Training and evaluation data

• shi3z/ja_conv_wikipedia_orion14B_100K

Training procedure

Training hyperparameters

The following hyperparameters were used during training:

• learning_rate: 0.0005
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 4
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 40
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss
2.684	0.06	40	2.7430
2.5302	0.11	80	2.6693
2.486	0.17	120	2.6273
2.557	0.23	160	2.6020
2.4913	0.28	200	2.5868
2.5317	0.34	240	2.5646
2.4795	0.4	280	2.5521
2.4221	0.45	320	2.5359
2.4464	0.51	360	2.5231
2.4534	0.57	400	2.5095
2.4685	0.62	440	2.4967
2.4575	0.68	480	2.4849
2.4299	0.74	520	2.4771
2.459	0.79	560	2.4604
2.4585	0.85	600	2.4527
2.4832	0.91	640	2.4425
2.4255	0.96	680	2.4285
2.2209	1.02	720	2.4312
2.3142	1.07	760	2.4288
2.1961	1.13	800	2.4252
2.1394	1.19	840	2.4194
2.2005	1.24	880	2.4093
2.0748	1.3	920	2.4003
2.109	1.36	960	2.3935
2.2209	1.41	1000	2.3856
2.1938	1.47	1040	2.3786
2.1056	1.53	1080	2.3716
2.0948	1.58	1120	2.3674
2.218	1.64	1160	2.3629
2.17	1.7	1200	2.3601
2.1084	1.75	1240	2.3590
2.0446	1.81	1280	2.3567
2.1517	1.87	1320	2.3572
2.2342	1.92	1360	2.3565
2.1552	1.98	1400	2.3564

Framework versions

• Transformers 4.34.1
• Pytorch 2.1.2
• Datasets 2.16.1
• Tokenizers 0.14.1