zephyr-7b-dpo-full-ultrabin-reward-scale-1-rpo

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.0345
• Rewards/chosen: -0.1240
• Rewards/rejected: -0.4158
• Rewards/accuracies: 0.7734
• Rewards/margins: 0.2918
• Logps/rejected: -304.2409
• Logps/chosen: -275.0308
• Logits/rejected: -2.4195
• Logits/chosen: -2.5003

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
0.0497	0.1046	50	0.0471	0.0341	-0.0639	0.6953	0.0980	-269.0505	-259.2180	-2.5747	-2.6118
0.0399	0.2092	100	0.0400	-0.0674	-0.3039	0.7656	0.2365	-293.0492	-269.3653	-2.2263	-2.2796
0.0384	0.3138	150	0.0368	-0.1521	-0.4051	0.7812	0.2530	-303.1761	-277.8396	-2.4575	-2.5017
0.0354	0.4184	200	0.0368	-0.1608	-0.4413	0.7812	0.2805	-306.7949	-278.7134	-2.6355	-2.6785
0.035	0.5230	250	0.0359	-0.0276	-0.3002	0.7812	0.2726	-292.6817	-265.3905	-2.5364	-2.5931
0.0336	0.6276	300	0.0351	-0.1609	-0.4489	0.7734	0.2880	-307.5566	-278.7195	-2.3179	-2.4060
0.0338	0.7322	350	0.0348	-0.1145	-0.3940	0.7695	0.2795	-302.0604	-274.0787	-2.3603	-2.4329
0.0352	0.8368	400	0.0345	-0.1250	-0.4112	0.7734	0.2863	-303.7862	-275.1277	-2.4372	-2.5111
0.0342	0.9414	450	0.0345	-0.1240	-0.4158	0.7734	0.2918	-304.2409	-275.0308	-2.4195	-2.5003

Framework versions

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