metadata
license: llama3.2
This is an experimental orpo finetune of llama 3.2 1b using our autoredteaming method, in addition to helpful instructions.
{
"cache_dir": "/leonardo_work/EUHPC_E03_068/.cache",
"method": "orpo",
"dataset": "autoredteam",
"lr": 0.01,
"train_batch_size": 5,
"eval_batch_size": 5,
"num_epochs": 3,
"seed": 42,
"eval_only": false,
"evaluation_size": 2000,
"checkpoint_path": null,
"experiment_group": "results",
"context_length": 1024,
"train_summarization": "",
"dpo_beta": 0.1,
"orpo_beta": 0.1,
"kl_coef": 0.0,
"reward_model": "",
"bestofn_size": 4,
"train_reward_model": ""
}