htlou/mm-interp-AA_preference_l0_new_step10_0_50

AA_preference_l0_new_step10_0_50

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_l0_new_step10_0_50 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5718
• Rewards/chosen: 1.6486
• Rewards/rejected: -0.6959
• Rewards/accuracies: 0.7875
• Rewards/margins: 2.3445
• Logps/rejected: -234.6941
• Logps/chosen: -260.0429
• Logits/rejected: -2.1761
• Logits/chosen: -2.2166

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
0.6007	0.7463	50	0.6151	1.2505	-0.1629	0.7625	1.4133	-229.3638	-264.0242	-2.0851	-2.1190
0.2744	1.4925	100	0.5826	1.7733	-0.3014	0.7542	2.0746	-230.7489	-258.7962	-2.2145	-2.2505
0.1721	2.2388	150	0.5628	1.4346	-0.7617	0.7833	2.1963	-235.3523	-262.1832	-2.1546	-2.1956
0.1926	2.9851	200	0.5716	1.6470	-0.6954	0.7875	2.3424	-234.6892	-260.0590	-2.1760	-2.2166

Framework versions

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