llama3-L1-SFT-L2-DPO

This model is a fine-tuned version of EllieS/TempReason-L1-llama3 on the EllieS/Temp-L2-DPO dataset. It achieves the following results on the evaluation set:

• Loss: 0.0028
• Rewards/chosen: -0.4016
• Rewards/rejected: -9.3312
• Rewards/accuracies: 1.0
• Rewards/margins: 8.9296
• Logps/rejected: -994.2072
• Logps/chosen: -84.5141
• Logits/rejected: 1.5698
• Logits/chosen: 0.6541

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-06
• train_batch_size: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• gradient_accumulation_steps: 4
• total_train_batch_size: 4
• 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
0.0033	0.2497	1000	0.0065	-0.3610	-7.6953	1.0	7.3344	-830.6234	-80.4518	1.5489	0.7452
0.0013	0.4995	2000	0.0031	-0.3798	-9.1892	1.0	8.8094	-980.0131	-82.3365	1.5546	0.6455
0.0019	0.7492	3000	0.0028	-0.3966	-9.3440	1.0	8.9474	-995.4902	-84.0208	1.5703	0.6568
0.0011	0.9989	4000	0.0028	-0.4016	-9.3312	1.0	8.9296	-994.2072	-84.5141	1.5698	0.6541

Framework versions

• PEFT 0.7.1
• Transformers 4.40.2
• Pytorch 2.1.2+cu121
• Datasets 2.18.0
• Tokenizers 0.19.1