zephyr-dpo-qlora-uf-ours-5e-6-epoch1

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the generation/UF dataset. It achieves the following results on the evaluation set:

• Loss: 0.9016
• Rewards/chosen: -5.0415
• Rewards/rejected: -5.9633
• Rewards/accuracies: 0.6560
• Rewards/margins: 0.9219
• Rewards/margins Max: 4.5250
• Rewards/margins Min: -2.8411
• Rewards/margins Std: 2.5079
• Logps/rejected: -854.9131
• Logps/chosen: -788.7391
• Logits/rejected: -1.3435
• Logits/chosen: -1.4046

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-06
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 2
• total_train_batch_size: 16
• total_eval_batch_size: 16
• 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	Rewards/margins Max	Rewards/margins Min	Rewards/margins Std	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.4419	0.28	100	0.6650	-0.4529	-0.5977	0.6240	0.1449	0.8619	-0.5418	0.4692	-318.3518	-329.8811	-2.5449	-2.5719
0.1872	0.56	200	0.7859	-2.8392	-3.5400	0.6630	0.7008	3.4374	-2.1827	1.9158	-612.5828	-568.5178	-1.4159	-1.4771
0.1102	0.85	300	0.8935	-4.8886	-5.8224	0.6470	0.9339	4.5089	-2.8033	2.4986	-840.8231	-773.4503	-1.3417	-1.4023

Framework versions

• PEFT 0.7.1
• Transformers 4.39.0.dev0
• Pytorch 2.1.2+cu121
• Datasets 2.14.6
• Tokenizers 0.15.2