ELYZA-japanese-Llama-2-7b-instruct-pun-2024-09-24-2024-09-28-22-26-11

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: 0.7531
• Rewards/chosen: 1.1327
• Logps/chosen: -367.4309
• Rewards/rejected: 1.0472
• Logps/rejected: -367.6175
• Rewards/margins: 0.0855
• Kl: 10.3562
• Logits/chosen: 8989739.0
• Logits/rejected: 8964266.0

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: 5e-07
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• total_train_batch_size: 64
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Logps/chosen	Rewards/rejected	Logps/rejected	Rewards/margins	Kl	Logits/chosen	Logits/rejected
0.7083	1.0	1095	0.7747	-0.1494	-380.2523	-0.1745	-379.8349	0.0250	0.0	9422631.0	9397082.0
0.671	2.0	2190	0.7541	0.8958	-369.7995	0.8179	-369.9113	0.0780	8.0609	9111132.0	9084323.0
0.6525	3.0	3285	0.7531	1.1327	-367.4309	1.0472	-367.6175	0.0855	10.3562	8989739.0	8964266.0

Framework versions

• PEFT 0.8.1
• Transformers 4.44.2
• Pytorch 2.4.0+cu121
• Datasets 2.18.0
• Tokenizers 0.19.1