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

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: 0.0010
• Rewards/chosen: -2.2051
• Rewards/rejected: -74.7846
• Rewards/accuracies: 1.0
• Rewards/margins: 72.5795
• Logps/rejected: -8119.2456
• Logps/chosen: -587.4921
• Logits/rejected: 2.7746
• Logits/chosen: -0.3615

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
0.0135	0.1420	50	0.0099	-1.3212	-47.4111	0.9960	46.0899	-5381.9028	-499.1044	-2.4386	-2.9291
0.0109	0.2841	100	0.0025	-2.5970	-80.9884	1.0	78.3913	-8739.6260	-626.6862	-0.0340	-2.5428
0.0017	0.4261	150	0.0011	-1.9943	-77.1591	1.0	75.1648	-8356.6973	-566.4090	2.8304	-1.6476
0.002	0.5682	200	0.0008	-2.1292	-82.8472	1.0	80.7180	-8925.5107	-579.9021	2.8840	-1.3436
0.0018	0.7102	250	0.0009	-2.1417	-72.3491	1.0	70.2074	-7875.6992	-581.1540	2.5447	-1.1532
0.0013	0.8523	300	0.0010	-2.2050	-73.9724	1.0	71.7674	-8038.0322	-587.4813	2.7313	-0.4348
0.0058	0.9943	350	0.0010	-2.2051	-74.7846	1.0	72.5795	-8119.2456	-587.4921	2.7746	-0.3615

Framework versions

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