zephyr-7b-mypo3_sim-full-beta10.0-lr4e-7

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: 1.3702
• Rewards/chosen: 0.0877
• Rewards/rejected: -0.3857
• Rewards/accuracies: 0.7480
• Rewards/margins: 0.4734
• Logps/rejected: -1.1662
• Logps/chosen: -0.9702
• Logits/rejected: -2.5842
• Logits/chosen: -2.6366

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: 4e-07
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• 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: 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
1.3832	0.1047	100	1.3867	0.1162	-0.1491	0.7361	0.2653	-1.1426	-0.9673	-2.5828	-2.6375
1.4139	0.2094	200	1.4360	-0.0655	-0.4534	0.7063	0.3879	-1.1730	-0.9855	-2.6262	-2.6797
1.3957	0.3141	300	1.3946	0.0975	-0.3093	0.7202	0.4069	-1.1586	-0.9692	-2.6662	-2.7142
1.376	0.4187	400	1.3812	0.1541	-0.2826	0.7242	0.4366	-1.1559	-0.9635	-2.5936	-2.6453
1.367	0.5234	500	1.3809	0.1133	-0.3635	0.7361	0.4767	-1.1640	-0.9676	-2.5923	-2.6449
1.3654	0.6281	600	1.3839	0.1611	-0.2773	0.7460	0.4384	-1.1554	-0.9628	-2.6239	-2.6742
1.3722	0.7328	700	1.3789	0.1067	-0.3804	0.7381	0.4870	-1.1657	-0.9683	-2.5906	-2.6428
1.3544	0.8375	800	1.3720	0.0906	-0.3825	0.7480	0.4731	-1.1659	-0.9699	-2.5825	-2.6349
1.3566	0.9422	900	1.3701	0.0867	-0.3862	0.7480	0.4729	-1.1663	-0.9703	-2.5839	-2.6363

Framework versions

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