biomistral-7b-dpo-full-wo-live_qa-ep3

This model is a fine-tuned version of BioMistral/BioMistral-7B on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5347
• Rewards/chosen: -0.8760
• Rewards/rejected: -1.7973
• Rewards/accuracies: 0.6528
• Rewards/margins: 0.9213
• Logps/rejected: -374.4584
• Logps/chosen: -385.7039
• Logits/rejected: 1.0068
• Logits/chosen: -1.3925

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: 4
• gradient_accumulation_steps: 2
• total_train_batch_size: 64
• total_eval_batch_size: 32
• 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	Logits/chosen	Logits/rejected	Logps/chosen	Logps/rejected	Validation Loss	Rewards/accuracies	Rewards/chosen	Rewards/margins	Rewards/rejected
0.3672	0.33	100	-1.7892	0.0124	-325.3138	-259.1306	0.5652	0.75	-0.2721	0.3720	-0.6441
0.2495	0.65	200	-1.4424	0.8731	-366.8614	-347.8152	0.5144	0.7153	-0.6876	0.8433	-1.5309
0.1708	0.98	300	-1.3915	1.0056	-385.5208	-374.2370	0.5345	0.6528	-0.8742	0.9209	-1.7951

Framework versions

• Transformers 4.39.0.dev0
• Pytorch 2.1.2
• Datasets 2.14.6
• Tokenizers 0.15.2