zephyr-7b-mypo3-qlora-beta1.5-lr5e6

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

• Loss: 1.3504
• Rewards/chosen: -0.0069
• Rewards/rejected: -0.3040
• Rewards/accuracies: 0.7383
• Rewards/margins: 0.2972
• Logps/rejected: -1.2481
• Logps/chosen: -0.8956
• Logits/rejected: -2.0523
• Logits/chosen: -2.1353

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: 8
• eval_batch_size: 16
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_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	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
1.3592	0.2093	100	1.3672	-0.0603	-0.2777	0.7227	0.2174	-1.2306	-0.9312	-2.0539	-2.1385
1.3573	0.4186	200	1.3550	-0.0161	-0.2742	0.7168	0.2582	-1.2282	-0.9018	-2.0756	-2.1575
1.3473	0.6279	300	1.3519	-0.0140	-0.3028	0.7422	0.2887	-1.2473	-0.9004	-2.0512	-2.1351
1.353	0.8373	400	1.3506	-0.0164	-0.3219	0.7402	0.3055	-1.2600	-0.9020	-2.0479	-2.1313

Framework versions

• PEFT 0.10.0
• Transformers 4.43.1
• Pytorch 2.1.2+cu121
• Datasets 2.18.0
• Tokenizers 0.19.1