lambda-qwen2.5-14b-dpo-test

This model is a fine-tuned version of Qwen/Qwen2.5-14B-Instruct on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.4919
• Rewards/chosen: -2.4745
• Rewards/rejected: -3.3729
• Rewards/accuracies: 0.7400
• Rewards/margins: 0.8984
• Logps/rejected: -832.0724
• Logps/chosen: -737.5234
• Logits/rejected: -1.2739
• Logits/chosen: -1.2560

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: 2
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 16
• 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
0.5269	0.2094	100	0.5333	-1.6756	-2.3320	0.7000	0.6564	-727.9815	-657.6356	-1.3952	-1.3850
0.5086	0.4187	200	0.5044	-2.0906	-2.9287	0.7040	0.8381	-787.6511	-699.1298	-1.2939	-1.2773
0.4787	0.6281	300	0.4948	-2.2927	-3.1689	0.7320	0.8762	-811.6696	-719.3386	-1.2846	-1.2646
0.4825	0.8375	400	0.4924	-2.4470	-3.3410	0.7400	0.8939	-828.8748	-734.7765	-1.2644	-1.2477

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1