MedQA_L3_300steps_1e6rate_05beta_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: 0.5731
• Rewards/chosen: 5.8136
• Rewards/rejected: 3.5872
• Rewards/accuracies: 0.7692
• Rewards/margins: 2.2264
• Logps/rejected: -26.6804
• Logps/chosen: -19.7013
• Logits/rejected: -0.8355
• Logits/chosen: -0.8339

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
0.6909	0.0489	50	0.6059	-0.4307	-0.6542	0.7538	0.2235	-35.1631	-32.1898	-0.7254	-0.7246
0.4343	0.0977	100	0.7202	6.9486	5.2431	0.6989	1.7054	-23.3686	-17.4314	-0.7816	-0.7804
0.6114	0.1466	150	0.6428	3.8385	1.9433	0.7407	1.8951	-29.9682	-23.6516	-0.8244	-0.8232
0.3522	0.1954	200	0.5948	5.1038	2.7837	0.7604	2.3201	-28.2874	-21.1208	-0.8383	-0.8367
0.3837	0.2443	250	0.5746	5.7825	3.5643	0.7692	2.2182	-26.7263	-19.7636	-0.8356	-0.8340
0.3658	0.2931	300	0.5731	5.8136	3.5872	0.7692	2.2264	-26.6804	-19.7013	-0.8355	-0.8339

Framework versions

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