tanliboy/lambda-gemma-2-9b-dpo

zephyr-gemma-2-9b-dpo-2

This model is a fine-tuned version of tanliboy/zephyr-gemma-2-9b-sft on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5628
• Rewards/chosen: -0.7292
• Rewards/rejected: -1.2825
• Rewards/accuracies: 0.6960
• Rewards/margins: 0.5533
• Logps/rejected: -1566.9301
• Logps/chosen: -1043.5624
• Logits/rejected: -14.1720
• Logits/chosen: -14.6638

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: 16
• total_train_batch_size: 256
• 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: 0.2
• 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
0.6835	0.2094	50	0.6815	-0.0218	-0.0436	0.6560	0.0218	-328.0053	-336.0947	-11.6381	-11.3403
0.6243	0.4187	100	0.6229	-0.5238	-0.7528	0.6600	0.2290	-1037.2136	-838.1255	-15.5098	-15.6787
0.5625	0.6281	150	0.5793	-0.7186	-1.1873	0.6880	0.4688	-1471.7362	-1032.8834	-14.7746	-15.1797
0.5699	0.8375	200	0.5647	-0.6443	-1.1499	0.6920	0.5057	-1434.3335	-958.5825	-14.1861	-14.6684

Framework versions

• Transformers 4.43.1
• Pytorch 2.3.1+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1