zephyr-7b-dpo-full-ultrabin-low-bleu-3-epochs

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: 0.5378
• Rewards/chosen: -1.1280
• Rewards/rejected: -2.2088
• Rewards/accuracies: 0.75
• Rewards/margins: 1.0808
• Logps/rejected: -483.5383
• Logps/chosen: -375.4253
• Logits/rejected: 0.4263
• Logits/chosen: -0.3049

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: 55
• 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: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6148	0.3484	50	0.6349	-0.0077	-0.1918	0.7344	0.1841	-281.8412	-263.4028	-2.4711	-2.5174
0.48	0.6969	100	0.5650	-0.3580	-0.9260	0.7383	0.5681	-355.2653	-298.4250	-0.5952	-0.8597
0.3945	1.0453	150	0.5507	-0.4890	-1.2040	0.7812	0.7150	-383.0621	-311.5342	0.6723	0.1059
0.3213	1.3937	200	0.5293	-0.7390	-1.5264	0.7617	0.7874	-415.3053	-336.5349	0.7764	0.2111
0.3246	1.7422	250	0.5303	-0.7632	-1.6866	0.7695	0.9235	-431.3257	-338.9464	0.1592	-0.5392
0.1986	2.0906	300	0.5372	-0.9618	-1.9577	0.7578	0.9959	-458.4319	-358.8135	0.6117	-0.1931
0.1848	2.4390	350	0.5348	-1.0734	-2.1515	0.7578	1.0781	-477.8110	-369.9702	0.4674	-0.2725
0.1901	2.7875	400	0.5385	-1.1315	-2.2179	0.7461	1.0864	-484.4519	-375.7825	0.4592	-0.2833

Framework versions

• Transformers 4.44.0.dev0
• Pytorch 2.1.2
• Datasets 2.20.0
• Tokenizers 0.19.1