num_input_tokens_seen value is set to 0

Hello, I've been training a model using fastchat/train/train_with_template.py, but when I checked the trainer_state.json file in the checkpoint directory, I noticed that the "num_input_tokens_seen" value is set to 0. Does anyone know why this value might be zero?

Here's a snippet from the log:

{
  "epoch": 0.36,
  "learning_rate": 1.4870847744984537e-05,
  "loss": 0.0778,
  "step": 4499
},
{
  "epoch": 0.36,
  "learning_rate": 1.4868531607943422e-05,
  "loss": 0.0757,
  "step": 4500
},
{
  "epoch": 0.36,
  "eval_loss": 0.07724634557962418,
  "eval_runtime": 3034.3515,
  "eval_samples_per_second": 42.499,
  "eval_steps_per_second": 2.656,
  "step": 4500
},
"logging_steps": 1.0,
"max_steps": 12341,
"num_input_tokens_seen": 0,
"num_train_epochs": 1,
"save_steps": 1500,
"total_flos": 2.3383019045781504e+19,
"train_batch_size": 4,
"trial_name": null,
"trial_params": null

I'm curious about the cause of this issue and how it might affect the training process.

Start with fastchat/train/train_with_template.py and inspect how training state is produced for the checkpoint directory. Trace where trainer_state.json gets its num_input_tokens_seen value, then compare the recorded state with the training and evaluation logs to determine whether zero is expected or indicates missing tracking; done means identifying the cause and documenting or correcting the behavior.