Llama-2-7b-hf-DPO-LookAhead-5_Q2_TTree1.4_TT0.9_TP0.7_TE0.2_V4

This model is a fine-tuned version of meta-llama/Llama-2-7b-hf on the None dataset. It achieves the following results on the evaluation set:

• Loss: 0.6849
• Rewards/chosen: -1.3255
• Rewards/rejected: -1.6674
• Rewards/accuracies: 0.5
• Rewards/margins: 0.3419
• Logps/rejected: -134.9012
• Logps/chosen: -95.5139
• Logits/rejected: 0.0033
• Logits/chosen: 0.1072

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-05
• train_batch_size: 2
• eval_batch_size: 2
• seed: 42
• gradient_accumulation_steps: 2
• total_train_batch_size: 4
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_steps: 10
• 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
0.7204	0.3043	63	0.6808	0.0801	0.0519	0.7000	0.0281	-117.7079	-81.4587	0.4903	0.5915
0.6989	0.6087	126	0.6930	0.0550	0.0726	0.6000	-0.0176	-117.5013	-81.7093	0.4748	0.5762
0.6896	0.9130	189	0.6579	0.1170	0.0536	0.5	0.0633	-117.6909	-81.0896	0.4569	0.5574
0.3332	1.2174	252	0.6831	-0.2141	-0.2394	0.5	0.0253	-120.6211	-84.4000	0.3842	0.4834
0.3687	1.5217	315	0.7069	-0.6436	-0.7406	0.5	0.0970	-125.6332	-88.6952	0.2816	0.3799
0.2083	1.8261	378	0.6389	-0.4156	-0.5567	0.5	0.1411	-123.7943	-86.4158	0.2329	0.3317
0.1191	2.1304	441	0.6451	-0.8600	-1.1248	0.5	0.2648	-129.4748	-90.8590	0.1067	0.2079
0.1435	2.4348	504	0.6878	-1.2620	-1.5788	0.5	0.3168	-134.0153	-94.8793	0.0284	0.1320
0.0848	2.7391	567	0.6849	-1.3255	-1.6674	0.5	0.3419	-134.9012	-95.5139	0.0033	0.1072

Framework versions

• PEFT 0.12.0
• Transformers 4.45.2
• Pytorch 2.4.0+cu121
• Datasets 3.2.0
• Tokenizers 0.20.3