zephyr-7b-align-scan-2e-07-0.5-linear-2.0

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

• Loss: 0.6464
• Rewards/chosen: 0.4833
• Rewards/rejected: -0.1353
• Rewards/accuracies: 0.3710
• Rewards/margins: 0.6186
• Logps/rejected: -81.3989
• Logps/chosen: -73.5245
• Logits/rejected: -2.5552
• Logits/chosen: -2.5716

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: 2e-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: linear
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 2

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6562	0.3484	100	0.6328	0.6802	0.3841	0.3552	0.2961	-80.3602	-73.1308	-2.5405	-2.5565
0.6601	0.6969	200	0.6410	0.2989	-0.0897	0.3452	0.3887	-81.3078	-73.8934	-2.5169	-2.5332
0.4195	1.0453	300	0.6371	0.6242	0.1593	0.3532	0.4648	-80.8097	-73.2429	-2.5193	-2.5354
0.3956	1.3937	400	0.6460	0.4324	-0.1472	0.3631	0.5796	-81.4227	-73.6264	-2.5378	-2.5541
0.3945	1.7422	500	0.6465	0.3072	-0.2969	0.3710	0.6040	-81.7221	-73.8769	-2.5543	-2.5709

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.0
• Datasets 2.21.0
• Tokenizers 0.19.1