htlou/mm-interp-AA_preference_cosi_new_step10_0_70

AA_preference_cosi_new_step10_0_70

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_cosi_new_step10_0_70 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5004
• Rewards/chosen: 0.6967
• Rewards/rejected: -2.0150
• Rewards/accuracies: 0.8006
• Rewards/margins: 2.7116
• Logps/rejected: -216.6696
• Logps/chosen: -251.2154
• Logits/rejected: -2.3392
• Logits/chosen: -2.3776

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
0.5778	0.5348	50	0.5711	1.2153	-0.0689	0.7560	1.2843	-197.2090	-246.0286	-2.3540	-2.3824
0.232	1.0695	100	0.4935	0.9389	-1.0786	0.8036	2.0175	-207.3059	-248.7935	-2.4977	-2.5262
0.2427	1.6043	150	0.4994	0.6147	-1.7279	0.8185	2.3426	-213.7988	-252.0355	-2.1963	-2.2464
0.1239	2.1390	200	0.4941	0.6670	-1.8992	0.8065	2.5662	-215.5122	-251.5120	-2.3605	-2.3958
0.1607	2.6738	250	0.5006	0.6972	-2.0217	0.8006	2.7190	-216.7369	-251.2096	-2.3349	-2.3732

Framework versions

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