410M_it-132366_Tr-slim-pajama-300B-replay5_finetune_dpo-full

This model is a fine-tuned version of btherien/410M_it-132366_Tr-slim-pajama-300B-replay5_finetune_sft-full on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.6500
• Rewards/chosen: -0.4023
• Rewards/rejected: -0.5547
• Rewards/accuracies: 0.6171
• Rewards/margins: 0.1514
• Logps/rejected: -420.0
• Logps/chosen: -464.0
• Logits/rejected: -1.3828
• Logits/chosen: -1.4297

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: 5e-07
• train_batch_size: 8
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 4
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_batch_size: 32
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6867	0.21	100	0.6853	-0.0056	-0.0271	0.5556	0.0216	-368.0	-424.0	-1.1953	-1.2422
0.6632	0.42	200	0.6625	-0.2285	-0.3203	0.625	0.0913	-396.0	-446.0	-1.3203	-1.3672
0.6493	0.63	300	0.6532	-0.3555	-0.4902	0.625	0.1348	-412.0	-460.0	-1.3516	-1.3984
0.6462	0.84	400	0.6495	-0.4004	-0.5547	0.6389	0.1553	-420.0	-464.0	-1.375	-1.4219

Framework versions

• Transformers 4.36.2
• Pytorch 2.0.1+cu117
• Datasets 2.14.6
• Tokenizers 0.15.2