zephyr-7b-align-scan-7e-07-0.99-cosine-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: 1.1727
• Rewards/chosen: 2.6978
• Rewards/rejected: 0.6816
• Rewards/accuracies: 0.3472
• Rewards/margins: 2.0163
• Logps/rejected: -80.4399
• Logps/chosen: -71.7662
• Logits/rejected: -2.6444
• Logits/chosen: -2.6613

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: 7e-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: cosine
• 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
0.9118	0.3484	100	0.8952	1.8796	1.1355	0.3353	0.7441	-79.9814	-72.5926	-2.5564	-2.5727
0.9553	0.6969	200	1.0700	2.5989	1.4006	0.3413	1.1983	-79.7136	-71.8661	-2.5726	-2.5893
0.4066	1.0453	300	1.0729	2.3164	0.9125	0.3433	1.4038	-80.2066	-72.1515	-2.5962	-2.6126
0.3805	1.3937	400	1.1546	2.9774	1.1937	0.3373	1.7837	-79.9225	-71.4837	-2.6247	-2.6413
0.3975	1.7422	500	1.1824	2.5463	0.5342	0.3452	2.0120	-80.5887	-71.9193	-2.6463	-2.6632

Framework versions

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