mistral-7b-dpo-full-sft-wo-live_qa

This model is a fine-tuned version of Minbyul/mistral-7b-wo-live_qa-sft on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.0469
• Rewards/chosen: -0.7784
• Rewards/rejected: -14.5261
• Rewards/accuracies: 0.875
• Rewards/margins: 13.7477
• Logps/rejected: -2053.0686
• Logps/chosen: -168.4323
• Logits/rejected: -1.5819
• Logits/chosen: -1.9164

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: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• 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
0.2894	0.3	100	0.2025	-0.2133	-2.0069	0.875	1.7935	-801.1411	-111.9223	-2.6944	-2.4093
0.1355	0.61	200	0.0506	-0.8364	-12.1329	0.875	11.2965	-1813.7421	-174.2286	-1.5654	-1.9005
0.0848	0.91	300	0.0468	-0.7745	-14.5125	0.875	13.7379	-2051.7014	-168.0429	-1.5809	-1.9182

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.1.2
• Datasets 2.14.6
• Tokenizers 0.15.2