zephyr-7b-uf-rlced-conifer-1e2e-group-dpo-2e

This model is a fine-tuned version of alignment-handbook/zephyr-7b-sft-full on the data/zephyr_uf_rlced_conifer_ref_1e2e dataset. It achieves the following results on the evaluation set:

• Loss: 0.2626
• Rewards/chosen: -2.1843
• Rewards/rejected: -5.4288
• Rewards/accuracies: 0.8684
• Rewards/margins: 3.2445
• Logps/rejected: -946.6157
• Logps/chosen: -610.9032
• Logits/rejected: 1.2318
• Logits/chosen: -0.7806
• Excess Loss: 0.0374
• Alpha 0 Uf: 0.8470
• Alpha 1 Rlced Conifer: 0.1530
• Rewards/chosen 1 Rlced Conifer: -2.2281
• Rewards/rejected 1 Rlced Conifer: -6.0246
• Rewards/accuracies 1 Rlced Conifer: 0.8987
• Rewards/margins 1 Rlced Conifer: 3.7965
• Logps/rejected 1 Rlced Conifer: -1049.9939
• Logps/chosen 1 Rlced Conifer: -646.3860
• Logits/rejected 1 Rlced Conifer: 1.1158
• Logits/chosen 1 Rlced Conifer: -0.9982
• Task Loss 1 Rlced Conifer: 0.2102
• Task Excess Loss 1 Rlced Conifer: 0.0475
• Rewards/chosen 0 Uf: -1.9978
• Rewards/rejected 0 Uf: -3.3091
• Rewards/accuracies 0 Uf: 0.7603
• Rewards/margins 0 Uf: 1.3113
• Logps/rejected 0 Uf: -572.5212
• Logps/chosen 0 Uf: -489.0419
• Logits/rejected 0 Uf: 1.8243
• Logits/chosen 0 Uf: -0.1004
• Task Loss 0 Uf: 0.4944
• Task Excess Loss 0 Uf: 0.0469

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: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 256
• 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: 2

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen	Excess Loss	Alpha 0 Uf	Alpha 1 Rlced Conifer	Rewards/chosen 1 Rlced Conifer	Rewards/rejected 1 Rlced Conifer	Rewards/accuracies 1 Rlced Conifer	Rewards/margins 1 Rlced Conifer	Logps/rejected 1 Rlced Conifer	Logps/chosen 1 Rlced Conifer	Logits/rejected 1 Rlced Conifer	Logits/chosen 1 Rlced Conifer	Task Loss 1 Rlced Conifer	Task Excess Loss 1 Rlced Conifer	Rewards/chosen 0 Uf	Rewards/rejected 0 Uf	Rewards/accuracies 0 Uf	Rewards/margins 0 Uf	Logps/rejected 0 Uf	Logps/chosen 0 Uf	Logits/rejected 0 Uf	Logits/chosen 0 Uf	Task Loss 0 Uf	Task Excess Loss 0 Uf
0.1953	0.4997	360	0.3535	-1.5938	-3.1996	0.8402	1.6058	-723.6984	-551.8521	0.1112	-0.7863	0.1136	0.9694	0.0306	-1.5989	-3.4179	0.8677	1.8190	-789.3262	-583.4747	-0.1145	-0.9516	0.3087	0.1414	-1.5520	-2.3972	0.7448	0.8452	-481.3242	-444.4588	1.0137	-0.2527	0.5289	0.0768
0.1537	0.9993	720	0.3329	-1.4289	-3.2979	0.8609	1.8690	-733.5210	-535.3586	0.6830	-0.5276	0.0943	0.9852	0.0148	-1.4038	-3.4887	0.8869	2.0849	-796.4048	-563.9600	0.3914	-0.7372	0.2955	0.1278	-1.4972	-2.5982	0.7618	1.1009	-501.4233	-438.9818	1.8477	0.1514	0.4804	0.0530
0.0667	1.4990	1080	0.2667	-2.1402	-5.1839	0.8656	3.0437	-922.1221	-606.4852	1.0002	-0.7884	0.0408	0.8954	0.1046	-2.1729	-5.7323	0.8964	3.5594	-1020.7665	-640.8754	0.8903	-0.9784	0.2150	0.0521	-1.9916	-3.2293	0.7574	1.2377	-564.5363	-488.4239	1.5582	-0.1961	0.4940	0.0466
0.06	1.9986	1440	0.2626	-2.1843	-5.4288	0.8684	3.2445	-946.6157	-610.9032	1.2318	-0.7806	0.0374	0.8470	0.1530	-2.2281	-6.0246	0.8987	3.7965	-1049.9939	-646.3860	1.1158	-0.9982	0.2102	0.0475	-1.9978	-3.3091	0.7603	1.3113	-572.5212	-489.0419	1.8243	-0.1004	0.4944	0.0469

Framework versions

• Transformers 4.44.1
• Pytorch 2.1.2+cu121
• Datasets 2.21.0
• Tokenizers 0.19.1