htlou/mm-interp-AA_preference_cosi_0_75

AA_preference_cosi_0_75

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_cosi_0_75 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5439
• Rewards/chosen: 1.0617
• Rewards/rejected: -1.0230
• Rewards/accuracies: 0.7292
• Rewards/margins: 2.0847
• Logps/rejected: -221.4094
• Logps/chosen: -257.3853
• Logits/rejected: -2.2898
• Logits/chosen: -2.2986

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
0.5811	0.7463	50	0.5699	0.5697	-0.6122	0.7333	1.1819	-217.3008	-262.3048	-2.3887	-2.3796
0.2898	1.4925	100	0.5633	1.2446	-0.5551	0.7583	1.7998	-216.7303	-255.5556	-2.4747	-2.4717
0.131	2.2388	150	0.5345	1.2941	-0.7142	0.7625	2.0083	-218.3207	-255.0607	-2.3181	-2.3241
0.1357	2.9851	200	0.5440	1.0620	-1.0267	0.7333	2.0886	-221.4456	-257.3822	-2.2899	-2.2988

Framework versions

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