[Bug] In generated UnslothSFTTrainer.py line 632

[sxndqc]

All previous checking happens before "skip_prepare_dataset", which causes problem because I didn't pass in any parameters for dataset_text_field. The direct returning of dataset should be put at the first lines.

[rolandtannous]

Hello @sxndqc please:

• provide information about your environment and package versions
• full code that would allows us to reproduce the error you've encountered so we can track it so we can verify that it's a bug and deal it with properly.
  Thank you

[Premik]

I think I'm hitting the same issue in unsloth v2025.3.19. When using the SFTTrainer/UnslothTrainer trainer.
I don't use the text field but have my data already tokenized in the token_ids and labels columns. So I set dataset_text_field=None and skip_prepare_dataset=True. But I got this:

 File ".../unsloth_compiled_cache/UnslothSFTTrainer.py", line 690, in _prepare_dataset
    test_text = dataset[0][dataset_text_field] if (formatting_func is None and dataset_text_field is not None) else formatting_func(dataset[0])[0]
                                                                                                                    ^^^^^^^^^^^^^^^^^^^^^^^^^^^
TypeError: 'NoneType' object is not callable

As sxndqc mentioned while there is a check for this flag in the original code, unsloth patch is inserted before this check.

A simple fix is to just repeat this check as part of the unsloth patch. I'll open a PR which should hopefully make what Imean clear.

This is a simple code to reproduce the issue:

import os
from datasets import Dataset
from unsloth import FastLanguageModel, UnslothTrainer, UnslothTrainingArguments
import torch

dataset = Dataset.from_dict(
    {
        # "text": "12",
        "input_ids": [[1, 2, 3, 0], [4, 5, 6, 0]],
        "attention_mask": [[1, 1, 1, 0], [1, 1, 1, 0]],
        "labels": [[1, 2, 3, -100], [4, 5, 6, -100]],
    }
)

model, tokenizer = FastLanguageModel.from_pretrained(model_name="unsloth/gemma-3-1b-it")
model = FastLanguageModel.get_peft_model(model)


args = UnslothTrainingArguments(
    per_device_train_batch_size=1,
    num_train_epochs=1,
    output_dir="/tmp/test_out",
    report_to="none",
    dataset_kwargs={"skip_prepare_dataset": True},  # This should skip ds processing
)


trainer = UnslothTrainer(
    model=model,
    processing_class=tokenizer,
    train_dataset=dataset,
    dataset_text_field=None,  # No text field since we have input_ids already
    max_seq_length=512,
    args=args,
)

trainer.train()

[mmathew23]

@Premik Does dataset_text_field = "" work?

[Premik]

@Premik Does dataset_text_field = "" work?

It only changes the error to:

  File "/wrk/ai/proj/text/unsloth/unsloth_compiled_cache/UnslothSFTTrainer.py", line 690, in _prepare_dataset
    test_text = dataset[0][dataset_text_field] if (formatting_func is None and dataset_text_field is not None) else formatting_func(dataset[0])[0]
                ~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^
KeyError: ''

The PR #2633

[github-actions]

Is this issue still important to you? Apologies in advance we might have missed this issue as well. For faster response times, please post on our Reddit server - https://www.reddit.com/r/unsloth or our Discord - https://discord.com/invite/unsloth