Mistral-7B-base-simpo-qlora_baseline

This model is a fine-tuned version of /scratch/wangxiaobo/zephyr-7b-sft-qlora on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 1.3519
• Rewards/chosen: -4.2809
• Rewards/rejected: -5.4373
• Rewards/accuracies: 0.7040
• Rewards/margins: 1.1564
• Logps/rejected: -2.7187
• Logps/chosen: -2.1404
• Logits/rejected: -1.6611
• Logits/chosen: -1.7141

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: 1e-05
• train_batch_size: 8
• eval_batch_size: 4
• seed: 42
• gradient_accumulation_steps: 8
• total_train_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	Logits/chosen	Logits/rejected	Logps/chosen	Logps/rejected	Validation Loss	Rewards/accuracies	Rewards/chosen	Rewards/margins	Rewards/rejected
1.3653	0.4187	400	-1.7568	-1.7004	-2.0068	-2.5244	1.4022	0.6910	-4.0135	1.0353	-5.0489
1.4162	0.8375	800	1.3523	-4.2715	-5.4265	0.7030	1.1551	-2.7133	-2.1357	-1.6619	-1.7153

Framework versions

• PEFT 0.11.1
• Transformers 4.42.2
• Pytorch 2.0.1+cu117
• Datasets 2.20.0
• Tokenizers 0.19.1