llama-3.2-3b-dpo

This model is a fine-tuned version of tanliboy/llama-3.2-3b-sft on the HuggingFaceH4/orca_dpo_pairs and the HuggingFaceH4/ultrafeedback_binarized datasets. It achieves the following results on the evaluation set:

• Loss: 0.6289
• Rewards/chosen: 0.7479
• Rewards/rejected: -3.8379
• Rewards/accuracies: 0.7405
• Rewards/margins: 4.5857
• Logps/rejected: -370.2327
• Logps/chosen: -338.3392
• Logits/rejected: 0.4475
• Logits/chosen: 0.3731

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: 1e-06
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• 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: 0.03
• num_epochs: 3

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.5801	0.4739	100	0.6840	0.6485	-2.9389	0.6899	3.5875	-361.2435	-339.3325	0.6783	0.6103
0.537	0.9479	200	0.6514	0.2045	-4.0315	0.7278	4.2360	-372.1696	-343.7731	0.5648	0.4948
0.4787	1.4218	300	0.6387	0.4099	-3.9882	0.7215	4.3981	-371.7361	-341.7187	0.5326	0.4589
0.4559	1.8957	400	0.6332	0.7690	-3.6688	0.7342	4.4379	-368.5425	-338.1277	0.4841	0.4110
0.4028	2.3697	500	0.6289	0.7479	-3.8379	0.7405	4.5857	-370.2327	-338.3392	0.4475	0.3731
0.4029	2.8436	600	0.6284	0.8504	-3.7058	0.7437	4.5562	-368.9125	-337.3143	0.4571	0.3820

Framework versions

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