zephyr-7b-dpo-full-gpt_consistent-reward-scale-1

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: 0.4815
• Rewards/chosen: -1.5849
• Rewards/rejected: -2.8045
• Rewards/accuracies: 0.7328
• Rewards/margins: 1.2196
• Logps/rejected: -526.9686
• Logps/chosen: -443.5758
• Logits/rejected: 3.4838
• Logits/chosen: 2.3333

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
0.6607	0.1147	50	0.6447	-0.0044	-0.1452	0.6897	0.1408	-261.0398	-285.5275	-2.4923	-2.5735
0.5616	0.2294	100	0.5464	-0.8527	-1.5772	0.6853	0.7245	-404.2408	-370.3625	0.2075	-0.2410
0.5333	0.3440	150	0.5195	-1.0024	-1.8820	0.7112	0.8797	-434.7274	-385.3255	1.4808	0.5890
0.5219	0.4587	200	0.5010	-1.0719	-2.0541	0.7328	0.9822	-451.9354	-392.2838	2.4260	1.4256
0.5007	0.5734	250	0.4917	-1.2321	-2.3291	0.7241	1.0970	-479.4298	-408.2994	2.6738	1.4527
0.5109	0.6881	300	0.4878	-1.3356	-2.5048	0.7284	1.1691	-496.9991	-418.6534	2.8884	1.5762
0.5063	0.8028	350	0.4814	-1.4870	-2.6833	0.7371	1.1963	-514.8549	-433.7904	3.3469	2.1699
0.4936	0.9174	400	0.4815	-1.5849	-2.8045	0.7328	1.2196	-526.9686	-443.5758	3.4838	2.3333

Framework versions

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