unexpected keyword argument tokenizer [FIXED]

[avemio-digital]

I used orpo colab example for mistral model and I am getting this error. I am using below configs

from trl import ORPOConfig, ORPOTrainer
from unsloth import is_bfloat16_supported

orpo_trainer = ORPOTrainer(
model = model,
train_dataset = dataset,
tokenizer = tokenizer,
args = ORPOConfig(
max_length = max_seq_length,
max_prompt_length = max_seq_length//2,
max_completion_length = max_seq_length//2,
per_device_train_batch_size = 2,
gradient_accumulation_steps = 4,
beta = 0.1,
logging_steps = 1,
optim = "adamw_8bit",
lr_scheduler_type = "linear",
max_steps = 1500, # Change to num_train_epochs = 1 for full training runs
fp16 = not is_bfloat16_supported(),
bf16 = is_bfloat16_supported(),
output_dir = "outputs",
report_to = "none", # Use this for WandB etc
),
)

[dame-cell]

So unsloth uses the newest version of trl always and just recently trl just updated and remove the tokenizer parameter to processing_class mentioned here 2348
you can simply do this

- trainer = DPOTrainer(model, args=training_args, train_dataset=dataset, tokenizer=tokenizer)
+ trainer = DPOTrainer(model, args=training_args, train_dataset=dataset, processing_class=tokenizer)

[Erland366]

Unsloth will make backward compability I think

[danielhanchen]

Just fixed @avemio-digital @dame-cell ! Please update Unsloth on local machines via pip install --upgrade --no-cache-dir --no-deps unsloth. For Colab, Kaggle, just refresh!

[qgallouedec]

This has been fixed in trl v0.12.1, make sure to run

pip install --upgrade trl

please also note that tokenizer is deprecated, so you should prefer processing_class, as mentioned by @dame-cell in #1285 (comment)