zephyr-7b-dpo-full-ultrabin-reward-scale-05

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: 0.5419
• Rewards/chosen: -2.0652
• Rewards/rejected: -3.5533
• Rewards/accuracies: 0.7812
• Rewards/margins: 1.4880
• Logps/rejected: -617.9895
• Logps/chosen: -469.1545
• Logits/rejected: 3.0313
• Logits/chosen: 2.1931

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: 55
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 2
• total_train_batch_size: 128
• total_eval_batch_size: 64
• 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
0.6746	0.1046	50	0.6514	0.0215	-0.0844	0.6953	0.1059	-271.1068	-260.4849	-2.5751	-2.6121
0.5801	0.2092	100	0.5963	-1.2413	-2.0024	0.6914	0.7611	-462.9021	-386.7607	0.8478	0.5614
0.561	0.3138	150	0.5612	-1.3516	-2.3053	0.7422	0.9537	-493.1910	-397.7852	2.1227	1.6750
0.552	0.4184	200	0.5634	-1.7910	-3.0147	0.7539	1.2237	-564.1274	-441.7259	2.6771	2.0183
0.5367	0.5230	250	0.5404	-1.6069	-2.8715	0.7656	1.2646	-549.8127	-423.3247	2.8098	2.1736
0.5231	0.6276	300	0.5511	-1.8243	-3.2523	0.7656	1.4280	-587.8877	-445.0558	2.9864	2.2075
0.5092	0.7322	350	0.5402	-1.9840	-3.4024	0.7734	1.4184	-602.9061	-461.0307	2.8834	2.0946
0.5231	0.8368	400	0.5417	-2.0950	-3.5645	0.7812	1.4695	-619.1116	-472.1271	3.0542	2.2365
0.5232	0.9414	450	0.5419	-2.0657	-3.5528	0.7812	1.4871	-617.9430	-469.2008	3.0322	2.1926

Framework versions

• Transformers 4.44.0.dev0
• Pytorch 2.1.2
• Datasets 2.20.0
• Tokenizers 0.19.1