htlou/mm-interp-RLAIF-V-Cosi-q0_25_preference

RLAIF-V-Cosi-q0_25_preference

This model is a fine-tuned version of llava-hf/llava-v1.6-mistral-7b-hf on the RLAIF-V-Cosi-q0_25_preference dataset. It achieves the following results on the evaluation set:

• Loss: 0.5909
• Rewards/chosen: -2.7988
• Rewards/rejected: -4.2628
• Rewards/accuracies: 0.7266
• Rewards/margins: 1.4639
• Logps/rejected: -211.2034
• Logps/chosen: -194.0678
• Logits/rejected: -2.6060
• Logits/chosen: -2.6047

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
0.575	0.6944	50	0.5680	-0.5232	-1.1499	0.6914	0.6267	-180.0744	-171.3114	-2.7966	-2.7982
0.2161	1.3889	100	0.5416	-1.1269	-2.2892	0.7461	1.1623	-191.4681	-177.3486	-2.6708	-2.6714
0.0912	2.0833	150	0.5559	-2.1342	-3.5698	0.7188	1.4356	-204.2739	-187.4216	-2.6701	-2.6674
0.0828	2.7778	200	0.5902	-2.7717	-4.2295	0.7227	1.4578	-210.8705	-193.7959	-2.6071	-2.6057

Framework versions

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