zephyr-7b-align-scan-7e-07-0.45-cosine-3.0

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: 0.9379
• Rewards/chosen: -0.3417
• Rewards/rejected: -2.0404
• Rewards/accuracies: 0.3472
• Rewards/margins: 1.6986
• Logps/rejected: -85.6625
• Logps/chosen: -75.2506
• Logits/rejected: -2.6727
• Logits/chosen: -2.6887

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: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6948	0.3484	100	0.6997	0.9816	0.4942	0.3452	0.4873	-80.0301	-72.3100	-2.5413	-2.5577
0.7373	0.6969	200	0.7720	1.2732	0.5117	0.3294	0.7615	-79.9912	-71.6619	-2.5716	-2.5870
0.4002	1.0453	300	0.8163	0.4524	-0.4497	0.3472	0.9021	-82.1276	-73.4859	-2.6256	-2.6409
0.3982	1.3937	400	0.8872	1.2165	0.0680	0.3313	1.1485	-80.9772	-71.7879	-2.7106	-2.7265
0.389	1.7422	500	0.9107	0.3181	-0.9594	0.3353	1.2775	-83.2604	-73.7844	-2.7188	-2.7346
0.3707	2.0906	600	0.8992	0.6908	-0.7854	0.3472	1.4762	-82.8736	-72.9561	-2.6904	-2.7065
0.3672	2.4390	700	0.9354	-0.5110	-2.2396	0.3492	1.7285	-86.1051	-75.6269	-2.6662	-2.6823
0.3596	2.7875	800	0.9344	-0.3373	-2.0235	0.3452	1.6862	-85.6249	-75.2407	-2.6727	-2.6886

Framework versions

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