htlou/mm-interp-AA_preference_cocour_new_step10_0_70

AA_preference_cocour_new_step10_0_70

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_cocour_new_step10_0_70 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5221
• Rewards/chosen: 0.8511
• Rewards/rejected: -1.7632
• Rewards/accuracies: 0.7827
• Rewards/margins: 2.6143
• Logps/rejected: -233.5941
• Logps/chosen: -246.8883
• Logits/rejected: -2.2402
• Logits/chosen: -2.2674

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
0.5857	0.5348	50	0.5588	0.8626	-0.4673	0.7202	1.3299	-220.6345	-246.7726	-2.3950	-2.4070
0.2515	1.0695	100	0.5002	1.1633	-0.7568	0.7827	1.9201	-223.5298	-243.7656	-2.3802	-2.3927
0.2749	1.6043	150	0.5163	0.8806	-1.4497	0.7738	2.3303	-230.4583	-246.5927	-2.4269	-2.4377
0.148	2.1390	200	0.5244	1.0476	-1.4591	0.7887	2.5067	-230.5529	-244.9229	-2.2151	-2.2425
0.1437	2.6738	250	0.5224	0.8628	-1.7441	0.7798	2.6069	-233.4026	-246.7709	-2.2343	-2.2617

Framework versions

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