smollm2-1.7B-8k-mix7-ep2-v2-qlora-r16-a16-lr3e4-mix1-dpo

This model is a fine-tuned version of gabrielmbmb/smollm2-1.7B-8k-mix7-ep2-v2-qlora-r16-a16-lr3e4-mix1 on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.5156
• Rewards/chosen: -4.3948
• Rewards/rejected: -5.4305
• Rewards/accuracies: 0.7656
• Rewards/margins: 1.0357
• Logps/rejected: -849.2994
• Logps/chosen: -752.2209
• Logits/rejected: -1.2614
• Logits/chosen: -1.2341

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: 0.0003
• train_batch_size: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_batch_size: 64
• 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
0.5629	0.2093	100	0.5765	-2.2363	-2.7515	0.7227	0.5152	-581.3958	-536.3666	-1.6836	-1.6572
0.5499	0.4186	200	0.5310	-3.6681	-4.4704	0.7734	0.8022	-753.2846	-679.5535	-0.8209	-0.8334
0.511	0.6279	300	0.5225	-4.8153	-5.7685	0.7695	0.9532	-883.0991	-794.2732	-1.1394	-1.1157
0.508	0.8373	400	0.5157	-4.3456	-5.3593	0.7695	1.0137	-842.1772	-747.2964	-1.2109	-1.1871

Framework versions

• PEFT 0.13.2
• Transformers 4.45.0
• Pytorch 2.4.1+cu121
• Datasets 3.0.2
• Tokenizers 0.20.1