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

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

• Loss: 0.7233
• Rewards/chosen: -3.4588
• Rewards/rejected: -4.7204
• Rewards/accuracies: 0.7148
• Rewards/margins: 1.2616
• Logps/rejected: -734.7018
• Logps/chosen: -608.5129
• Logits/rejected: 1.4610
• Logits/chosen: 1.1037

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
0.6745	0.3484	50	0.6653	-0.0371	-0.1111	0.6484	0.0740	-273.7737	-266.3441	-2.5668	-2.6013
0.6331	0.6969	100	0.6233	-0.4237	-0.7043	0.6836	0.2806	-333.0932	-305.0008	-2.2639	-2.3200
0.521	1.0453	150	0.5987	-0.6985	-1.1642	0.7383	0.4657	-379.0843	-332.4787	-2.0284	-2.1059
0.4172	1.3937	200	0.6243	-1.9881	-2.8118	0.7461	0.8237	-543.8417	-461.4368	0.1952	-0.0218
0.3987	1.7422	250	0.6016	-2.1537	-3.0839	0.7656	0.9301	-571.0483	-478.0042	0.2787	-0.0206
0.2489	2.0906	300	0.6069	-2.2987	-3.3106	0.7539	1.0119	-593.7270	-492.5032	0.4389	0.1163
0.2214	2.4390	350	0.7216	-3.3725	-4.6253	0.7109	1.2527	-725.1878	-599.8825	1.3133	0.9733
0.2176	2.7875	400	0.7233	-3.4588	-4.7204	0.7148	1.2616	-734.7018	-608.5129	1.4610	1.1037

Framework versions

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