htlou/mm-interp-AA_preference_cocour_new_step10_0_80

AA_preference_cocour_new_step10_0_80

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the AA_preference_cocour_new_step10_0_80 dataset. It achieves the following results on the evaluation set:

• Loss: 0.5250
• Rewards/chosen: 0.5658
• Rewards/rejected: -2.2697
• Rewards/accuracies: 0.8307
• Rewards/margins: 2.8355
• Logps/rejected: -237.9764
• Logps/chosen: -250.5795
• Logits/rejected: -2.1447
• Logits/chosen: -2.1680

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
0.572	0.4673	50	0.5720	1.0708	-0.2185	0.7578	1.2893	-217.4643	-245.5292	-2.5291	-2.5309
0.4997	0.9346	100	0.5102	0.8653	-0.9135	0.7865	1.7788	-224.4143	-247.5850	-2.1776	-2.2008
0.2873	1.4019	150	0.5675	1.0559	-1.2279	0.7891	2.2838	-227.5579	-245.6786	-2.2632	-2.2750
0.2853	1.8692	200	0.5163	0.7188	-1.7114	0.8203	2.4302	-232.3931	-249.0491	-2.1251	-2.1478
0.1541	2.3364	250	0.5271	0.5977	-2.1434	0.8177	2.7411	-236.7135	-250.2604	-2.2153	-2.2352
0.1566	2.8037	300	0.5242	0.5568	-2.2821	0.8307	2.8389	-238.1007	-250.6694	-2.1442	-2.1674

Framework versions

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