Llama0-3-8b-ultra-p-0.05

This model is a fine-tuned version of meta-llama/Meta-Llama-3-8B-Instruct on an unknown dataset. It achieves the following results on the evaluation set:

• Loss: 0.5203
• Rewards/chosen: -0.9516
• Rewards/rejected: -1.8713
• Rewards/accuracies: 0.7344
• Rewards/margins: 0.9197
• Logps/rejected: -451.7943
• Logps/chosen: -351.7126
• Logits/rejected: 0.7305
• Logits/chosen: 0.5934

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: 2
• eval_batch_size: 8
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 128
• total_eval_batch_size: 64
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: linear
• num_epochs: 2.0

Training results

Training Loss	Epoch	Step	Validation Loss	Rewards/chosen	Rewards/rejected	Rewards/accuracies	Rewards/margins	Logps/rejected	Logps/chosen	Logits/rejected	Logits/chosen
0.6104	0.2060	100	0.6071	-0.3701	-0.6034	0.6719	0.2333	-325.0017	-293.5632	0.3051	0.2437
0.5839	0.4119	200	0.5788	-0.4556	-0.8422	0.6875	0.3867	-348.8867	-302.1120	0.3006	0.2201
0.561	0.6179	300	0.5644	-0.5556	-1.0628	0.6953	0.5072	-370.9420	-312.1160	0.4520	0.3342
0.5489	0.8239	400	0.5474	-0.6136	-1.2310	0.7188	0.6174	-387.7600	-317.9102	0.5350	0.4135
0.5112	1.0299	500	0.5325	-0.8101	-1.5835	0.7344	0.7734	-423.0114	-337.5664	0.6096	0.4735
0.4577	1.2358	600	0.5291	-1.0553	-1.9636	0.7109	0.9083	-461.0272	-362.0868	0.7787	0.6411
0.4688	1.4418	700	0.5220	-0.9496	-1.8535	0.7266	0.9040	-450.0144	-351.5097	0.7421	0.6045
0.4684	1.6478	800	0.5194	-0.9894	-1.9115	0.7344	0.9221	-455.8119	-355.4933	0.7759	0.6367
0.4731	1.8538	900	0.5202	-0.9713	-1.9015	0.7344	0.9303	-454.8165	-353.6810	0.7319	0.5955

Framework versions

• Transformers 4.45.1
• Pytorch 2.4.1+cu121
• Datasets 3.0.0
• Tokenizers 0.20.0