Fix SFTTrainer TypeError: Unexpected Keyword Argument 'tokenizer'

This error occurs when initializing SFTTrainer from the trl library with the tokenizer argument, which was renamed to processing_class in trl version 0.12.0. The exact error message is:

TypeError: SFTTrainer.__init__() got an unexpected keyword argument 'tokenizer'

This is a breaking change introduced in the 0.12.0 release of trl. The most common cause is using an older code pattern with a newer version of the library.

What Causes This Error

According to the accepted answer on Stack Overflow (source 2), the error has one primary cause:

1. API change in trl 0.12.0: The tokenizer parameter was renamed to processing_class. This is documented in the trl release notes for version 0.12.0. If you are using trl >= 0.12.0 and your code passes tokenizer=tokenizer to SFTTrainer, you will get this TypeError because the old parameter name no longer exists.

2. Version mismatch: If you are running in an environment where trl was automatically updated (e.g., Google Colab, which is mentioned in the question from source 1), you may have unknowingly upgraded to a version that removed the old parameter. The user in source 1 was running their script in Google Colab, which often has pre-installed packages that get updated over time.

3. Incorrect parameter passing: The user in source 1 also tried passing tokenizer inside training_arguments (which is an SFTConfig object), but that did not work. The SFTConfig class does not accept a tokenizer parameter; it has its own set of arguments like dataset_text_field and max_seq_length. The tokenizer must be passed directly to SFTTrainer, but with the new name.

How to Fix It

Solution 1: Replace tokenizer with processing_class (Recommended)

This is the fix provided in the accepted answer by nemo on Stack Overflow (source 2). In trl 0.12.0 and later, the tokenizer argument was renamed to processing_class. You should change your SFTTrainer initialization to use the new parameter name.

Step 1: Update your SFTTrainer call

Change this:

trainer = SFTTrainer(
    model=model,
    train_dataset=data,
    peft_config=peft_config,
    args=training_arguments,
    tokenizer=tokenizer,  # This causes the error
)

To this:

trainer = SFTTrainer(
    model=model,
    train_dataset=data,
    peft_config=peft_config,
    args=training_arguments,
    processing_class=tokenizer,  # Fixed: use processing_class instead of tokenizer
)

Step 2: Verify the fix

After making this change, the TypeError should no longer occur. The processing_class parameter accepts the same tokenizer object you were passing before. In the user's code from source 1, the tokenizer was loaded with:

tokenizer = AutoTokenizer.from_pretrained("TheBloke/Mistral-7B-Instruct-v0.1-GPTQ")
tokenizer.pad_token = tokenizer.eos_token

This tokenizer object can be passed directly to processing_class. The SFTTrainer will use it for tokenizing the dataset during training.

Step 3: Ensure other parameters are correct

The user in source 1 also had packing=False and max_seq_length=512 in their SFTConfig. These are still valid parameters in trl 0.12.0. The dataset_text_field="text" parameter tells SFTTrainer which column in the dataset contains the text to tokenize. Make sure your dataset has a column with that name (the user created it with a lambda function that combined instruction, input, and output into a single "text" field).

Solution 2: Downgrade trl to a version that supports tokenizer

If you cannot or do not want to change your code, you can downgrade trl to a version before 0.12.0. This is a community-reported workaround (source 1 implies the user was unaware of the version change, and source 2 confirms the API change).

Step 1: Check your current trl version

Run this command in your environment:

pip show trl

Look for the Version line. If it is 0.12.0 or higher, you are affected.

Step 2: Downgrade to a compatible version

Install a version that still supports the tokenizer parameter. For example:

pip install trl==0.11.0

Note: The exact version that last supported tokenizer may vary. Check the trl release notes on GitHub to find the version where the change was introduced. According to source 2, the change was made in 0.12.0, so versions 0.11.x and earlier should work.

Step 3: Restart your kernel or runtime

After downgrading, restart your Python kernel (in a notebook) or your script. The error should no longer appear because the old parameter name is recognized again.

Caveat: Downgrading may cause other issues if you rely on features added in 0.12.0. This is a temporary workaround, not a long-term solution.

Solution 3: Use SFTConfig correctly (if you were trying to pass tokenizer there)

The user in source 1 mentioned trying to pass tokenizer inside training_arguments. This is not a valid approach. The SFTConfig class does not accept a tokenizer parameter. Instead, SFTConfig has parameters like dataset_text_field, max_seq_length, and packing that control how the dataset is processed. The tokenizer must be passed to SFTTrainer directly via processing_class (Solution 1) or via the old tokenizer parameter (if you downgrade).

If you were attempting to pass the tokenizer through SFTConfig, remove that attempt and use Solution 1 instead.