htlou/mm-interp-AA_preference_cosi_new_step10_0_50

AA_preference_cosi_new_step10_0_50

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_cosi_new_step10_0_50 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5631
• Rewards/chosen: 1.1577
• Rewards/rejected: -1.6637
• Rewards/accuracies: 0.8208
• Rewards/margins: 2.8214
• Logps/rejected: -228.8032
• Logps/chosen: -253.1133
• Logits/rejected: -2.2359
• Logits/chosen: -2.2654

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
0.5242	0.7463	50	0.5642	0.8117	-0.7280	0.7583	1.5397	-219.4454	-256.5733	-2.1958	-2.2182
0.2523	1.4925	100	0.5878	1.3931	-1.2040	0.8042	2.5970	-224.2055	-250.7594	-2.3529	-2.3702
0.1264	2.2388	150	0.5602	1.3270	-1.4230	0.8167	2.7499	-226.3953	-251.4204	-2.2382	-2.2652
0.1301	2.9851	200	0.5625	1.1589	-1.6695	0.8208	2.8283	-228.8604	-253.1016	-2.2358	-2.2653

Framework versions

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