lambda-llama-3-8b-ipo-test

This model is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B-Instruct on the HuggingFaceH4/ultrafeedback_binarized and the tanliboy/orca_dpo_pairs datasets. It achieves the following results on the evaluation set:

• Loss: 0.8931
• Rewards/chosen: -0.3610
• Rewards/rejected: -0.5883
• Rewards/accuracies: 0.7922
• Rewards/margins: 0.2272
• Logps/rejected: -3.1373
• Logps/chosen: -2.5334
• Logits/rejected: -2.9939
• Logits/chosen: -2.9244

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-07
• train_batch_size: 4
• eval_batch_size: 4
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 4
• total_train_batch_size: 128
• total_eval_batch_size: 32
• 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
1.1749	0.1744	100	1.0763	-0.1732	-0.3120	0.7892	0.1388	-2.4465	-2.0638	-2.5676	-2.5133
0.9802	0.3489	200	0.9501	-0.3184	-0.5302	0.8012	0.2118	-2.9922	-2.4269	-2.7873	-2.7230
0.9548	0.5233	300	0.9136	-0.3761	-0.6028	0.8163	0.2267	-3.1736	-2.5710	-2.8788	-2.8087
0.9834	0.6978	400	0.9041	-0.3384	-0.5537	0.8042	0.2153	-3.0509	-2.4770	-2.9371	-2.8667
0.9967	0.8722	500	0.8938	-0.3750	-0.6076	0.7892	0.2326	-3.1855	-2.5684	-3.0293	-2.9592

Framework versions

• Transformers 4.44.2
• Pytorch 2.4.0+cu121
• Datasets 2.19.1
• Tokenizers 0.19.1