This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

Loss: 0.1119
Rewards/chosen: -0.5472
Rewards/rejected: -1.1549
Rewards/accuracies: 0.75
Rewards/margins: 0.6076
Logps/rejected: -375.6880
Logps/chosen: -336.6998
Logits/rejected: -0.4187
Logits/chosen: -0.8032

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: 4
gradient_accumulation_steps: 4
total_train_batch_size: 128
total_eval_batch_size: 32
optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
lr_scheduler_type: cosine
lr_scheduler_warmup_ratio: 0.1
num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.8666	0.2093	100	0.1307	-0.4641	-1.0037	0.7282	0.5395	-360.5680	-328.3875	-0.7152	-1.0291
2.5975	0.4186	200	0.1221	-0.5115	-1.1127	0.7440	0.6012	-371.4751	-333.1310	-0.1960	-0.5634
0.0974	0.6279	300	0.1175	-0.5182	-1.0980	0.7540	0.5798	-370.0028	-333.7931	-0.6215	-0.9932
0.0828	0.8373	400	0.1121	-0.5476	-1.1540	0.7480	0.6065	-375.6061	-336.7349	-0.4274	-0.8105

Framework versions

Transformers 4.44.2
Pytorch 2.2.1+cu118
Datasets 2.14.7
Tokenizers 0.19.1

YeongminKim/zephyr-7b-dpo-full-alpha_0.5_batch128

zephyr-7b-dpo-full-alpha_0.5_batch128

Model description

Intended uses & limitations

Training and evaluation data

Training procedure

Training hyperparameters

Training results

Framework versions

YeongminKim/
zephyr-7b-dpo-full-alpha_0.5_batch128