htlou/mm-interp-AA_preference_cosi_0_50

AA_preference_cosi_0_50

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_cosi_0_50 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5580
• Rewards/chosen: 1.1191
• Rewards/rejected: -1.7161
• Rewards/accuracies: 0.8125
• Rewards/margins: 2.8352
• Logps/rejected: -229.3270
• Logps/chosen: -253.4991
• Logits/rejected: -2.2872
• Logits/chosen: -2.3146

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
0.5248	0.7463	50	0.5636	0.7981	-0.7376	0.7625	1.5357	-219.5417	-256.7093	-2.1893	-2.2123
0.2521	1.4925	100	0.5887	1.3477	-1.2511	0.8042	2.5988	-224.6767	-251.2128	-2.3607	-2.3766
0.1273	2.2388	150	0.5529	1.1944	-1.5829	0.8250	2.7773	-227.9948	-252.7459	-2.2852	-2.3106
0.1289	2.9851	200	0.5586	1.1189	-1.7195	0.8125	2.8384	-229.3605	-253.5011	-2.2866	-2.3140

Framework versions

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