htlou/mm-interp-AA_preference_random_0_90

AA_preference_random_0_90

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_random_0_90 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5266
• Rewards/chosen: 0.6411
• Rewards/rejected: -1.9030
• Rewards/accuracies: 0.7986
• Rewards/margins: 2.5441
• Logps/rejected: -230.4333
• Logps/chosen: -239.3183
• Logits/rejected: -2.0706
• Logits/chosen: -2.1025

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
0.635	0.4158	50	0.5987	0.8451	-0.0388	0.7014	0.8840	-211.7915	-237.2780	-2.3908	-2.3899
0.4933	0.8316	100	0.5285	-0.2263	-1.8151	0.7523	1.5888	-229.5545	-247.9923	-1.9128	-1.9530
0.2495	1.2474	150	0.5427	0.5572	-1.4201	0.7593	1.9773	-225.6041	-240.1570	-2.0983	-2.1232
0.2753	1.6632	200	0.5260	0.5776	-1.6735	0.7870	2.2511	-228.1382	-239.9529	-1.9752	-2.0068
0.1584	2.0790	250	0.5118	0.5255	-1.9057	0.7940	2.4312	-230.4605	-240.4746	-2.0354	-2.0689
0.1572	2.4948	300	0.5261	0.7582	-1.7260	0.7986	2.4842	-228.6629	-238.1469	-2.0616	-2.0941
0.1557	2.9106	350	0.5265	0.6414	-1.9061	0.7986	2.5475	-230.4645	-239.3154	-2.0706	-2.1026

Framework versions

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