htlou/mm-interp-AA_preference_l0_new_step10_0_70

AA_preference_l0_new_step10_0_70

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_l0_new_step10_0_70 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5015
• Rewards/chosen: 0.2932
• Rewards/rejected: -2.5932
• Rewards/accuracies: 0.8274
• Rewards/margins: 2.8864
• Logps/rejected: -244.7975
• Logps/chosen: -248.1062
• Logits/rejected: -2.1491
• Logits/chosen: -2.2057

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
0.5372	0.5348	50	0.5589	0.8885	-0.7687	0.7708	1.6573	-226.5533	-242.1535	-2.5558	-2.5652
0.2535	1.0695	100	0.5072	0.6618	-1.4593	0.8155	2.1211	-233.4590	-244.4205	-2.3160	-2.3489
0.2927	1.6043	150	0.5305	0.2199	-2.4527	0.8065	2.6726	-243.3929	-248.8394	-2.2493	-2.2896
0.1722	2.1390	200	0.4972	0.5435	-2.1567	0.8304	2.7003	-240.4332	-245.6031	-2.1267	-2.1841
0.1412	2.6738	250	0.5014	0.2961	-2.5802	0.8214	2.8763	-244.6681	-248.0778	-2.1488	-2.2053

Framework versions

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