simpo-oh-dcft-v3.1-llama-3.1-405b

This model is a fine-tuned version of mlfoundations-dev/oh-dcft-v3.1-llama-3.1-405b on the mlfoundations-dev/gemma2-ultrafeedback-armorm dataset. It achieves the following results on the evaluation set:

• Loss: 2.5125
• Rewards/chosen: -34.0628
• Rewards/rejected: -41.8860
• Rewards/accuracies: 0.8018
• Rewards/margins: 7.8232
• Logps/chosen: -3.4063
• Logps/rejected: -4.1886
• Logits/chosen: -1.4278
• Logits/rejected: -1.4170

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: 8e-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: Use OptimizerNames.ADAMW_TORCH with betas=(0.9,0.999) and epsilon=1e-08 and optimizer_args=No additional optimizer arguments
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 1.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/chosen	Logps/rejected	Logits/chosen	Logits/rejected
2.8265	0.9997	442	2.5125	-34.0628	-41.8860	0.8018	7.8232	-3.4063	-4.1886	-1.4278	-1.4170

Framework versions

• Transformers 4.46.1
• Pytorch 2.3.0
• Datasets 3.1.0
• Tokenizers 0.20.3