zephyr-7b-dpo-full

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: 0.5025
• Rewards/chosen: -1.0211
• Rewards/rejected: -2.0056
• Rewards/accuracies: 0.7773
• Rewards/margins: 0.9845
• Logps/rejected: -463.2262
• Logps/chosen: -364.7427
• Logits/rejected: 1.5124
• Logits/chosen: 0.6094

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
• gradient_accumulation_steps: 2
• 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
0.5655	0.2092	100	0.5746	-0.7161	-1.2966	0.7070	0.5805	-392.3193	-334.2372	-0.4365	-0.6501
0.5457	0.4184	200	0.5249	-0.9842	-1.7929	0.7773	0.8086	-441.9472	-361.0536	0.9213	0.3160
0.4955	0.6276	300	0.5092	-0.9649	-1.8656	0.7656	0.9007	-449.2220	-359.1224	1.2341	0.4705
0.5033	0.8368	400	0.5032	-1.0014	-1.9606	0.7734	0.9592	-458.7184	-362.7651	1.3908	0.5145

Framework versions

• Transformers 4.44.0
• Pytorch 2.4.0+cu121
• Datasets 2.21.0
• Tokenizers 0.19.1