Llama0-3-8b-ultra-p-0.025

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: 0.5336
• Rewards/chosen: -0.9176
• Rewards/rejected: -1.7954
• Rewards/accuracies: 0.7266
• Rewards/margins: 0.8778
• Logps/rejected: -444.2013
• Logps/chosen: -348.3104
• Logits/rejected: 0.6139
• Logits/chosen: 0.4767

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
0.6219	0.2060	100	0.6185	-0.3528	-0.5814	0.6797	0.2287	-322.8062	-291.8294	0.3035	0.2418
0.5957	0.4119	200	0.5909	-0.4280	-0.7949	0.6875	0.3669	-344.1528	-299.3566	0.2675	0.1885
0.5718	0.6179	300	0.5747	-0.5426	-1.0432	0.6797	0.5006	-368.9808	-310.8154	0.4283	0.3112
0.5604	0.8239	400	0.5587	-0.5860	-1.1683	0.7031	0.5824	-381.4986	-315.1523	0.4728	0.3510
0.5212	1.0299	500	0.5435	-0.7801	-1.5211	0.7344	0.7410	-416.7767	-334.5627	0.5275	0.3937
0.4671	1.2358	600	0.5421	-0.9911	-1.8546	0.7109	0.8635	-450.1224	-355.6647	0.6644	0.5299
0.4778	1.4418	700	0.5348	-0.9156	-1.7871	0.7266	0.8715	-443.3745	-348.1178	0.6300	0.4953
0.4791	1.6478	800	0.5330	-0.9374	-1.8128	0.7266	0.8754	-445.9473	-350.2950	0.6640	0.5278
0.4831	1.8538	900	0.5340	-0.9312	-1.8200	0.7266	0.8888	-446.6618	-349.6732	0.6213	0.4847

Framework versions

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