htlou/backup_0202_llamafactory_AA_preference_random_0_70-llava-mistral

AA_preference_random_0_70

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_random_0_70 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5414
• Rewards/chosen: 0.5872
• Rewards/rejected: -1.5508
• Rewards/accuracies: 0.7768
• Rewards/margins: 2.1380
• Logps/rejected: -227.1739
• Logps/chosen: -233.6734
• Logits/rejected: -1.9224
• Logits/chosen: -1.9691

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
0.6034	0.5348	50	0.5869	0.7400	-0.1413	0.7292	0.8813	-213.0786	-232.1449	-2.5305	-2.5372
0.2659	1.0695	100	0.5500	1.0886	-0.4196	0.7679	1.5081	-215.8615	-228.6596	-2.1146	-2.1482
0.2599	1.6043	150	0.5465	0.7596	-1.0841	0.7679	1.8437	-222.5069	-231.9491	-2.1122	-2.1442
0.1366	2.1390	200	0.5222	0.5890	-1.4225	0.7857	2.0115	-225.8904	-233.6554	-2.0824	-2.1154
0.1488	2.6738	250	0.5411	0.5980	-1.5260	0.7768	2.1240	-226.9253	-233.5653	-1.9322	-1.9781

Framework versions

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