A straightforward method for training your LLM, from downloading data to generating text.
{
"sft_ckpt": "models/sft.pt",
"reward_ckpt": "models/reward.pt",
"prompt_path": "data/rl_prompts_train.jsonl",
"eval_prompt_path": "data/rl_prompts_test.jsonl",
"out_ckpt": "models/ppo.pt",
"reward_source": "verifier",
"iterations": 1000,
"prompts_per_iter": 32,
"rollout_len": 300,
"temperature": 1.0,
"top_p": 1.0,
"ppo_epochs": 4,
"minibatch_size": 16,
"clip": 0.2,
"vf_clip": 0.2,
"vf_coef": 0.5,
"ent_coef": 0.0,
"gamma": 1.0,
"gae_lambda": 0.95,
"kl_coef": 0.05,
"lr": 1e-06,
"grad_clip": 1.0,
"eval_every": 50,
"save_every": 100
}