Chapter 12

[ritesh2014]

Below Code:

trainer = SFTTrainer(
model=model,
train_dataset=dataset,
dataset_text_field="text",
tokenizer=tokenizer,
args=training_arguments,
max_seq_length=512,

# Leave this out for regular SFT
peft_config=peft_config,

)

Error:
TypeError: SFTTrainer.init() got an unexpected keyword argument 'dataset_text_field'

Tried to resolve using Gemini, but the solutions kept throwing new errors and the code got messed up.
Has anyone resolved this?

[MaartenGr]

Where are you using the above code, in Colab? If so, can you use the versions of transformers/trl as shown in the requirements.txt?

[ritesh2014]

Using Colab.

So changed -
!pip install -q accelerate peft bitsandbytes transformers trl sentencepiece

To-
!pip install -q accelerate==0.31.0 peft==0.11.1 bitsandbytes==0.43.1 transformers==4.41.2 trl==0.9.4 sentencepiece==0.2.0

Had to provide versions to all the libs as additional errors were cropping up.
Now it works. When you can, pls change the notebook in Github to prevent this occurence for others.

[MaartenGr]

@ritesh2014 Thanks for testing the updated requirements! I finally had some time to update this and it should work now. Also, the reason why this took a bit longer is a nice one (cool update in a couple of hours 😉).

[amina-mardiyyah]

For future reference or anyone who bumps into this error on Colab, you could also replace transformers TrainingArguments with SFTConfig from SFT instead. Then define parameters as below:

training_arguments = SFTConfig(
    output_dir=output_dir,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=4,
    optim="paged_adamw_32bit",
    learning_rate=2e-4,
    lr_scheduler_type="cosine",
    num_train_epochs=1,
    logging_steps=10,
    max_seq_length = 512,
    fp16=True,
    gradient_checkpointing=True,
    dataset_text_field="text", 
)

Then:

trainer = SFTTrainer(
     model=model,
     train_dataset=dataset,
     tokenizer=tokenizer,
     args=training_arguments,
 )