Description

This model was trained as part of the Reinforcement Learning - 24 project at Peking University, focusing on [dpo].

Authors

• Ejafa Bassam
• Yaroslav Ponomarenko

phi-3-mini-128k-instruct-dpo-lr-5e-07

This model is a fine-tuned version of microsoft/Phi-3-mini-128k-instruct on the princeton-nlp/llama3-ultrafeedback dataset. It achieves the following results on the evaluation set:

• Loss: 0.6096
• Rewards/chosen: -1.0852
• Rewards/rejected: -1.4834
• Rewards/accuracies: 0.6976
• Rewards/margins: 0.3982
• Logps/rejected: -434.2651
• Logps/chosen: -403.4777
• Logits/rejected: 1.6861
• Logits/chosen: 1.6753

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: 2
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• 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
0.62	0.8549	400	0.6104	-1.0659	-1.4533	0.6976	0.3875	-433.6641	-403.0910	1.6821	1.6709

Framework versions

• Transformers 4.41.2
• Pytorch 2.3.1+cu121
• Datasets 2.20.0
• Tokenizers 0.19.1