zephyr-7b-align-scan-0.0-0.2-polynomial-3

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: 0.6601
• Rewards/chosen: -1.0432
• Rewards/rejected: -1.4686
• Rewards/accuracies: 0.3313
• Rewards/margins: 0.4254
• Logps/rejected: -89.9617
• Logps/chosen: -80.7659
• Logits/rejected: -2.4878
• Logits/chosen: -2.5051

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: 7.526744872300726e-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: 3

Training results

Training Loss	Epoch	Step	Logits/chosen	Logits/rejected	Logps/chosen	Logps/rejected	Validation Loss	Rewards/accuracies	Rewards/chosen	Rewards/margins	Rewards/rejected
0.5502	1.0417	100	-2.5290	-2.5124	-73.3173	-81.0099	0.6403	0.3254	0.2781	0.2500	0.0281
0.377	2.0833	200	-2.5251	-2.5078	-75.2992	-83.7654	0.6467	0.3313	-0.1914	0.4333	-0.6247

Framework versions

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