zephyr-7b-dpo-full-gpt-high-curriculum

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.5179
• Rewards/chosen: -0.8125
• Rewards/rejected: -1.5680
• Rewards/accuracies: 0.7241
• Rewards/margins: 0.7555
• Logps/rejected: -402.4431
• Logps/chosen: -365.2540
• Logits/rejected: 1.3741
• Logits/chosen: 0.3005

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: 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
0.6573	0.1147	50	0.6482	-0.0593	-0.1537	0.6422	0.0944	-261.0173	-289.9336	-2.4140	-2.5168
0.5517	0.2294	100	0.5831	-0.4867	-1.0049	0.6940	0.5182	-346.1422	-332.6784	-0.1634	-0.6358
0.5596	0.3440	150	0.5497	-0.4238	-1.0012	0.7241	0.5774	-345.7715	-326.3861	-0.2421	-1.0045
0.557	0.4587	200	0.5398	-0.7669	-1.4634	0.7328	0.6965	-391.9895	-360.6939	0.6306	-0.2433
0.5483	0.5734	250	0.5334	-0.9092	-1.6482	0.7371	0.7390	-410.4661	-374.9231	1.1694	0.1535
0.5338	0.6881	300	0.5227	-0.7072	-1.4506	0.7241	0.7434	-390.7057	-354.7213	1.1201	0.0530
0.5111	0.8028	350	0.5173	-0.7777	-1.5283	0.7284	0.7506	-398.4796	-361.7773	1.3205	0.2474
0.5185	0.9174	400	0.5179	-0.8125	-1.5680	0.7241	0.7555	-402.4431	-365.2540	1.3741	0.3005

Framework versions

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