MedQA_L3_300steps_1e6rate_01beta_CSFTDPO

This model is a fine-tuned version of tsavage68/MedQA_L3_1000steps_1e6rate_SFT on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.4661
• Rewards/chosen: 0.6273
• Rewards/rejected: -0.3771
• Rewards/accuracies: 0.7604
• Rewards/margins: 1.0045
• Logps/rejected: -37.6261
• Logps/chosen: -25.0552
• Logits/rejected: -0.8801
• Logits/chosen: -0.8780

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: 300

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6869	0.0489	50	0.6696	-0.2211	-0.2710	0.7253	0.0498	-36.5645	-33.5400	-0.7298	-0.7290
0.4779	0.0977	100	0.5887	1.4526	1.0417	0.6945	0.4109	-23.4374	-16.8024	-0.8047	-0.8036
0.5155	0.1466	150	0.4976	0.6394	-0.2000	0.7363	0.8394	-35.8551	-24.9343	-0.8636	-0.8617
0.4245	0.1954	200	0.4924	0.0477	-0.9077	0.7648	0.9554	-42.9321	-30.8513	-0.8783	-0.8762
0.4563	0.2443	250	0.4675	0.6549	-0.3364	0.7560	0.9913	-37.2189	-24.7791	-0.8807	-0.8786
0.3066	0.2931	300	0.4661	0.6273	-0.3771	0.7604	1.0045	-37.6261	-25.0552	-0.8801	-0.8780

Framework versions

• Transformers 4.41.1
• Pytorch 2.0.0+cu117
• Datasets 2.19.1
• Tokenizers 0.19.1