Meta-Llama-3-8B-Instruct-DPO-QLoRA

This model is a fine-tuned version of data/Meta-Llama-3-8B-Instruct-Merged on the HuggingFaceH4/ultrafeedback_binarized dataset. It achieves the following results on the evaluation set:

• Loss: 0.4785
• Rewards/chosen: -2.3087
• Rewards/rejected: -3.5097
• Rewards/accuracies: 0.7760
• Rewards/margins: 1.2010
• Logps/rejected: -604.2300
• Logps/chosen: -507.2661
• Logits/rejected: -0.8568
• Logits/chosen: -0.8381

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: 4
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 2
• gradient_accumulation_steps: 4
• total_train_batch_size: 32
• total_eval_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: 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
0.6865	0.0523	100	0.6857	0.0202	0.0036	0.6810	0.0166	-252.9014	-274.3707	-0.6048	-0.5953
0.5773	0.1047	200	0.5802	-0.5398	-0.9390	0.7080	0.3992	-347.1614	-330.3779	-0.9408	-0.9089
0.546	0.1570	300	0.5337	-0.9951	-1.7352	0.7370	0.7401	-426.7812	-375.9071	-1.0937	-1.0510
0.501	0.2094	400	0.5120	-1.8215	-2.7617	0.7530	0.9401	-529.4277	-458.5479	-1.1011	-1.0595
0.4525	0.2617	500	0.5090	-1.9857	-3.0848	0.7510	1.0991	-561.7446	-474.9624	-0.9430	-0.9134
0.508	0.3141	600	0.5005	-2.2106	-3.1511	0.7600	0.9405	-568.3763	-497.4550	-0.9955	-0.9626
0.4852	0.3664	700	0.5028	-1.3971	-2.4127	0.7770	1.0156	-494.5317	-416.1026	-0.9794	-0.9476
0.5474	0.4187	800	0.4966	-1.7948	-2.7637	0.7670	0.9689	-529.6284	-455.8714	-0.9115	-0.8851
0.5246	0.4711	900	0.4943	-1.5285	-2.5416	0.7660	1.0131	-507.4219	-429.2431	-0.8138	-0.7980
0.4635	0.5234	1000	0.4908	-2.8177	-4.0337	0.7630	1.2160	-656.6334	-558.1610	-0.8713	-0.8521
0.4856	0.5758	1100	0.4817	-2.3661	-3.4921	0.7720	1.1260	-602.4694	-512.9990	-0.8044	-0.7913
0.5013	0.6281	1200	0.4860	-2.1162	-3.2907	0.7720	1.1745	-582.3287	-488.0108	-0.7890	-0.7745
0.4497	0.6805	1300	0.4850	-2.4840	-3.7371	0.7730	1.2531	-626.9694	-524.7895	-0.8096	-0.7940
0.4734	0.7328	1400	0.4833	-2.1466	-3.3699	0.7740	1.2233	-590.2520	-491.0496	-0.8148	-0.7990
0.4482	0.7851	1500	0.4812	-2.5061	-3.7160	0.7760	1.2100	-624.8656	-527.0021	-0.8423	-0.8246
0.4982	0.8375	1600	0.4787	-2.2293	-3.3886	0.7770	1.1593	-592.1224	-499.3264	-0.8377	-0.8203
0.4594	0.8898	1700	0.4790	-2.3679	-3.5723	0.7730	1.2044	-610.4911	-513.1796	-0.8566	-0.8379
0.4551	0.9422	1800	0.4786	-2.3275	-3.5261	0.7730	1.1986	-605.8722	-509.1397	-0.8587	-0.8397
0.4605	0.9945	1900	0.4785	-2.3086	-3.5093	0.7740	1.2007	-604.1885	-507.2548	-0.8544	-0.8360

Framework versions

• PEFT 0.11.1
• Transformers 4.41.0
• Pytorch 2.3.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1