training-runs/sft-python-math-1b-v2/args.json

{
  "base_repo_id": "Daisuke675/scugnizz-1b",
  "base_ckpt": "training-runs/pretrain-1b/checkpoint_best.pt",
  "datasets": [
    "iamtarun/python_code_instructions_18k_alpaca",
    "open-r1/OpenR1-Math-220k"
  ],
  "dataset_weights": [
    0.6,
    0.4
  ],
  "split": "train",
  "max_samples_per_dataset": 25000,
  "val_size": 1000,
  "packing": true,
  "block_size": 1024,
  "batch_size": 1,
  "grad_accum": 16,
  "epochs": 1,
  "max_steps": 0,
  "lr": 1.5e-06,
  "min_lr": 3e-07,
  "warmup_steps": 100,
  "weight_decay": 0.05,
  "grad_clip": 1.0,
  "dropout": 0.05,
  "eval_interval": 250,
  "save_interval": 250,
  "log_interval": 10,
  "eval_batches": 30,
  "seed": 42,
  "dtype": "auto",
  "out_dir": "runs/sft-python-math-1b-v2",
  "hub_repo_id": "Daisuke675/scugnizz-1b",
  "hub_path": "training-runs/sft-python-math-1b-v2",
  "push": true,
  "push_every_save": true
}