zephyr-7b-align-scan

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: 0.5960
• Rewards/chosen: -0.6252
• Rewards/rejected: -1.0907
• Rewards/accuracies: 0.3690
• Rewards/margins: 0.4655
• Logps/rejected: -190.1945
• Logps/chosen: -137.0085
• Logits/rejected: 0.9319
• Logits/chosen: 0.4522

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-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: 1

Training results

Training Loss	Epoch	Step	Logits/chosen	Logits/rejected	Logps/chosen	Logps/rejected	Validation Loss	Rewards/accuracies	Rewards/chosen	Rewards/margins	Rewards/rejected
0.6719	0.1047	100	-2.3487	-2.3310	-73.2865	-85.4479	0.6687	0.3274	0.0120	0.0552	-0.0432
0.6488	0.2093	200	-1.2584	-1.2396	-102.5743	-130.3725	0.6348	0.3373	-0.2808	0.2116	-0.4924
0.6331	0.3140	300	-1.1873	-1.0320	-120.1307	-157.0977	0.6195	0.3452	-0.4564	0.3033	-0.7597
0.6321	0.4186	400	0.0335	0.3728	-146.9637	-190.2757	0.6099	0.3631	-0.7247	0.3667	-1.0915
0.6318	0.5233	500	2.6547	2.9545	-155.4930	-204.6371	0.6105	0.3552	-0.8100	0.4251	-1.2351
0.5978	0.6279	600	0.9606	1.4420	-147.8560	-199.5121	0.6015	0.3591	-0.7336	0.4502	-1.1838
0.6113	0.7326	700	1.1833	1.7188	-150.6854	-204.9195	0.5986	0.3651	-0.7619	0.4760	-1.2379
0.5885	0.8373	800	0.5613	1.0128	-141.6925	-192.4845	0.5974	0.3690	-0.6720	0.4415	-1.1136
0.595	0.9419	900	0.4326	0.9106	-136.2882	-189.5506	0.5958	0.3710	-0.6180	0.4663	-1.0842

Framework versions

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