prm_version3_full_hf

This model is a fine-tuned version of meta-llama/Meta-Llama-3.1-8B-Instruct on the prm_conversations_prm_version3_math+webinstructsub-mcq+webinstructsub-oe+apps+gsm_mix_ref_hf dataset. It achieves the following results on the evaluation set:

• Loss: 0.1166

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: 1
• eval_batch_size: 1
• seed: 42
• distributed_type: multi-GPU
• num_devices: 8
• gradient_accumulation_steps: 8
• total_train_batch_size: 64
• total_eval_batch_size: 8
• optimizer: Adam with betas=(0.9,0.999) and epsilon=1e-08
• lr_scheduler_type: cosine
• lr_scheduler_warmup_ratio: 0.05
• num_epochs: 1

Training results

Training Loss	Epoch	Step	Validation Loss
0.1961	0.0461	500	0.2069
0.192	0.0921	1000	0.1930
0.1963	0.1382	1500	0.1833
0.1701	0.1843	2000	0.1748
0.1647	0.2303	2500	0.1687
0.1507	0.2764	3000	0.1630
0.1421	0.3225	3500	0.1579
0.1403	0.3685	4000	0.1528
0.1557	0.4146	4500	0.1485
0.1536	0.4607	5000	0.1441
0.1344	0.5067	5500	0.1399
0.1195	0.5528	6000	0.1355
0.1209	0.5989	6500	0.1316
0.137	0.6450	7000	0.1284
0.117	0.6910	7500	0.1253
0.116	0.7371	8000	0.1228
0.1259	0.7832	8500	0.1206
0.1147	0.8292	9000	0.1187
0.1175	0.8753	9500	0.1175
0.1117	0.9214	10000	0.1168
0.1133	0.9674	10500	0.1166

Framework versions

• Transformers 4.45.0
• Pytorch 2.5.1+cu124
• Datasets 3.1.0
• Tokenizers 0.20.3