zephyr-7b-align-scan-6e-07-0.53-polynomial-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: 0.8595
• Rewards/chosen: 1.8502
• Rewards/rejected: 0.6070
• Rewards/accuracies: 0.3393
• Rewards/margins: 1.2433
• Logps/rejected: -79.9832
• Logps/chosen: -71.0003
• Logits/rejected: -2.6929
• Logits/chosen: -2.7082

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: 6e-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: polynomial
• 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
0.7136	0.3484	100	0.7109	1.4011	0.8625	0.3512	0.5386	-79.5010	-71.8476	-2.5458	-2.5618
0.7461	0.6969	200	0.7643	1.0640	0.3687	0.3274	0.6952	-80.4327	-72.4838	-2.5601	-2.5759
0.3949	1.0453	300	0.7875	0.2070	-0.6350	0.3472	0.8420	-82.3265	-74.1006	-2.6135	-2.6292
0.3838	1.3937	400	0.8714	0.4396	-0.7042	0.3294	1.1438	-82.4571	-73.6618	-2.6266	-2.6422
0.371	1.7422	500	0.8639	0.6923	-0.5434	0.3393	1.2357	-82.1536	-73.1851	-2.6910	-2.7068

Framework versions

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