Llama-3.1-8B-Instruct-SAA-700

This model is a fine-tuned version of meta-llama/Llama-3.1-8B-Instruct on the bct_non_cot_dpo_700 dataset. It achieves the following results on the evaluation set:

• Loss: 0.0846
• Rewards/chosen: -0.0062
• Rewards/rejected: -0.0635
• Rewards/accuracies: 0.8857
• Rewards/margins: 0.0573
• Logps/rejected: -0.6353
• Logps/chosen: -0.0623
• Logits/rejected: -0.4422
• Logits/chosen: -0.3590
• Sft Loss: 0.0098
• Odds Ratio Loss: 0.7473

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: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 8
• total_train_batch_size: 16
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.1
• num_epochs: 10.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen	Sft Loss	Odds Ratio Loss
1.3943	1.2698	50	1.1722	-0.1129	-0.1708	0.8714	0.0579	-1.7085	-1.1294	-0.5079	-0.3998	0.1369	10.3529
0.2668	2.5397	100	0.1508	-0.0126	-0.0716	0.8857	0.0590	-0.7158	-0.1261	-0.4935	-0.3919	0.0160	1.3479
0.1305	3.8095	150	0.0939	-0.0069	-0.0601	0.8857	0.0531	-0.6007	-0.0692	-0.4467	-0.3595	0.0109	0.8298
0.126	5.0794	200	0.0885	-0.0065	-0.0608	0.8857	0.0542	-0.6076	-0.0653	-0.4471	-0.3614	0.0103	0.7822
0.0881	6.3492	250	0.0876	-0.0064	-0.0617	0.8857	0.0553	-0.6175	-0.0642	-0.4433	-0.3588	0.0102	0.7739
0.1042	7.6190	300	0.0846	-0.0062	-0.0635	0.8857	0.0573	-0.6353	-0.0623	-0.4422	-0.3590	0.0098	0.7473
0.1405	8.8889	350	0.0853	-0.0063	-0.0644	0.8857	0.0581	-0.6435	-0.0627	-0.4405	-0.3572	0.0099	0.7540

Framework versions

• PEFT 0.12.0
• Transformers 4.45.2
• Pytorch 2.3.0
• Datasets 2.19.0
• Tokenizers 0.20.0