mpt_1000_STEPS_1e6_rate_01_beta_DPO

This model is a fine-tuned version of mosaicml/mpt-7b-instruct on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.6555
• Rewards/chosen: -0.9911
• Rewards/rejected: -1.1284
• Rewards/accuracies: 0.6220
• Rewards/margins: 0.1372
• Logps/rejected: -32.8413
• Logps/chosen: -30.7037
• Logits/rejected: 12.5582
• Logits/chosen: 12.5620

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: 1e-06
• train_batch_size: 2
• eval_batch_size: 1
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 100
• training_steps: 1000

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.7012	0.1	100	0.6878	0.0402	0.0262	0.5516	0.0140	-21.2953	-20.3903	14.1969	14.1998
0.6605	0.2	200	0.6893	0.1209	0.0818	0.5670	0.0391	-20.7398	-19.5837	13.0519	13.0548
0.657	0.29	300	0.6715	-0.4737	-0.5524	0.5758	0.0787	-27.0816	-25.5295	13.1844	13.1876
0.6934	0.39	400	0.6676	-0.8625	-0.9556	0.5934	0.0932	-31.1138	-29.4168	12.8462	12.8498
0.6891	0.49	500	0.6641	-1.0231	-1.1288	0.6088	0.1057	-32.8455	-31.0235	12.6874	12.6909
0.6492	0.59	600	0.6564	-0.9706	-1.0997	0.6462	0.1291	-32.5548	-30.4985	12.7748	12.7786
0.6512	0.68	700	0.6569	-0.9892	-1.1224	0.6220	0.1332	-32.7819	-30.6846	12.6401	12.6438
0.6687	0.78	800	0.6556	-0.9937	-1.1300	0.6330	0.1363	-32.8571	-30.7290	12.5528	12.5566
0.6668	0.88	900	0.6552	-0.9899	-1.1276	0.6308	0.1376	-32.8330	-30.6916	12.5557	12.5594
0.5867	0.98	1000	0.6555	-0.9911	-1.1284	0.6220	0.1372	-32.8413	-30.7037	12.5582	12.5620

Framework versions

• Transformers 4.39.1
• Pytorch 2.0.0+cu117
• Datasets 2.18.0
• Tokenizers 0.15.2