zephyr-7b-dpo-full-gpt_consistent-reward-scale-05

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: 0.4931
• Rewards/chosen: -1.6483
• Rewards/rejected: -2.9482
• Rewards/accuracies: 0.7414
• Rewards/margins: 1.2999
• Logps/rejected: -541.3424
• Logps/chosen: -449.9160
• Logits/rejected: 3.3584
• Logits/chosen: 2.0773

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
0.6653	0.1147	50	0.6513	-0.0038	-0.1422	0.6940	0.1384	-260.7392	-285.4682	-2.4923	-2.5731
0.5699	0.2294	100	0.5578	-0.8876	-1.6121	0.6853	0.7245	-407.7342	-373.8545	0.8771	0.3393
0.5416	0.3440	150	0.5320	-1.1447	-2.1157	0.7026	0.9709	-458.0881	-399.5647	2.1978	1.2442
0.5318	0.4587	200	0.5122	-1.1906	-2.2306	0.7284	1.0400	-469.5803	-404.1460	2.4483	1.3448
0.5178	0.5734	250	0.5029	-1.4402	-2.5615	0.7284	1.1212	-502.6709	-429.1149	2.5336	1.3215
0.519	0.6881	300	0.4985	-1.4880	-2.6823	0.7371	1.1943	-514.7540	-433.8906	2.6886	1.3895
0.5137	0.8028	350	0.4931	-1.6128	-2.8601	0.7328	1.2473	-532.5296	-446.3658	3.1580	1.8716
0.5033	0.9174	400	0.4931	-1.6483	-2.9482	0.7414	1.2999	-541.3424	-449.9160	3.3584	2.0773

Framework versions

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