htlou/mm-interp-AA_preference_cocour_0_50

AA_preference_cocour_0_50

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_cocour_0_50 dataset. It achieves the following results on the evaluation set:

• Loss: 0.4988
• Rewards/chosen: 0.9568
• Rewards/rejected: -1.9780
• Rewards/accuracies: 0.8500
• Rewards/margins: 2.9348
• Logps/rejected: -217.8485
• Logps/chosen: -263.8239
• Logits/rejected: -2.1050
• Logits/chosen: -2.1590

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
0.6027	0.7463	50	0.5491	1.3803	-0.2983	0.8208	1.6786	-201.0510	-259.5887	-2.4356	-2.4575
0.2795	1.4925	100	0.5112	1.1590	-1.5093	0.8417	2.6683	-213.1614	-261.8016	-1.9102	-1.9756
0.1557	2.2388	150	0.5033	1.3754	-1.3325	0.8583	2.7079	-211.3931	-259.6372	-2.1170	-2.1696
0.1338	2.9851	200	0.4983	0.9563	-1.9762	0.8500	2.9325	-217.8308	-263.8291	-2.1047	-2.1588

Framework versions

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