zephyr-7b-dpo-full-gpt-low-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: 0.5229
• Rewards/chosen: -0.8152
• Rewards/rejected: -1.5392
• Rewards/accuracies: 0.7069
• Rewards/margins: 0.7241
• Logps/rejected: -399.5724
• Logps/chosen: -365.5233
• Logits/rejected: 1.4072
• Logits/chosen: 0.3892

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
0.6558	0.1147	50	0.6455	0.0044	-0.0965	0.6810	0.1009	-255.3015	-283.5690	-2.4863	-2.5882
0.5907	0.2294	100	0.5894	-0.2321	-0.5376	0.7069	0.3055	-299.4117	-307.2200	-2.4655	-2.5910
0.5657	0.3440	150	0.5474	-0.5168	-1.0293	0.7198	0.5125	-348.5750	-335.6879	-0.6546	-1.0350
0.5303	0.4587	200	0.5414	-1.0659	-1.7181	0.75	0.6522	-417.4532	-390.5937	0.7246	0.0707
0.5472	0.5734	250	0.5268	-0.8095	-1.4718	0.7155	0.6623	-392.8294	-364.9606	1.2657	0.4213
0.5517	0.6881	300	0.5284	-0.8914	-1.6145	0.7112	0.7231	-407.0940	-373.1438	1.3137	0.2994
0.4943	0.8028	350	0.5237	-0.8328	-1.5668	0.7112	0.7339	-402.3227	-367.2895	1.4252	0.4044
0.5335	0.9174	400	0.5229	-0.8152	-1.5392	0.7069	0.7241	-399.5724	-365.5233	1.4072	0.3892

Framework versions

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