dpo-grad-acc-16-train_filtered_full

This model is a fine-tuned version of google/gemma-2-9b-it on the jz666/gemma2-ultrafeedback-ppl-split dataset. It achieves the following results on the evaluation set:

• Loss: 0.4532
• Rewards/chosen: -2.7523
• Rewards/rejected: -3.7825
• Rewards/accuracies: 0.7705
• Rewards/margins: 1.0302
• Logps/rejected: -562.3438
• Logps/chosen: -463.9622
• Logits/rejected: -19.6359
• Logits/chosen: -19.6258

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: 4
• gradient_accumulation_steps: 16
• 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
0.486	0.8724	400	0.4543	-2.7276	-3.7515	0.7705	1.0239	-559.2502	-461.4991	-19.5816	-19.5679

Framework versions

• Transformers 4.44.2
• Pytorch 2.7.0+cu128
• Datasets 2.18.0
• Tokenizers 0.19.1