zephyr-7b-dpo-full-accumulation4

This model is a fine-tuned version of data/zephyr-7b-sft-full-accumulation2 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5032
• Rewards/chosen: -0.9893
• Rewards/rejected: -2.0234
• Rewards/accuracies: 0.7812
• Rewards/margins: 1.0341
• Logps/rejected: -462.7061
• Logps/chosen: -358.6745
• Logits/rejected: 3.3182
• Logits/chosen: 2.7991

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: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• 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: 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
0.59	0.2093	100	0.5946	-0.2826	-0.6651	0.7266	0.3825	-326.8777	-288.0025	-2.2764	-2.3187
0.5622	0.4186	200	0.5490	-0.5914	-1.2367	0.7578	0.6452	-384.0357	-318.8896	-1.6885	-1.7635
0.5069	0.6279	300	0.5186	-0.9189	-1.8568	0.7773	0.9379	-446.0468	-351.6352	3.7286	3.1924
0.5183	0.8373	400	0.5042	-1.0384	-2.0520	0.7773	1.0136	-465.5701	-363.5876	3.4727	2.9519

Framework versions

• Transformers 4.40.0
• Pytorch 2.1.2+cu121
• Datasets 2.19.0
• Tokenizers 0.19.1