htlou/backup_0202_llamafactory_AA_preference_Cherry_0_70-llava-mistral

AA_preference_Cherry_0_70

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_Cherry_0_70 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4557
• Rewards/chosen: 1.5899
• Rewards/rejected: -1.4032
• Rewards/accuracies: 0.8198
• Rewards/margins: 2.9930
• Logps/rejected: -237.9070
• Logps/chosen: -288.5520
• Logits/rejected: -2.0398
• Logits/chosen: -2.0633

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: 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_steps: 10
• num_epochs: 3.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.1988	1.0554	100	0.4581	1.2441	-0.7508	0.8052	1.9949	-231.3840	-292.0099	-2.0274	-2.0496
0.0898	2.1108	200	0.4581	1.2188	-1.4863	0.8285	2.7050	-238.7381	-292.2634	-2.0389	-2.0563

Framework versions

• Transformers 4.45.2
• Pytorch 2.4.0+cu121
• Datasets 2.21.0
• Tokenizers 0.20.3