TRL importation error on Kaggle

Beginners
1

Hello. Please does anyone know if there is an issue with TRL library or a way to fix it? I am trying to fine tune Mistral 7B inside Kaggle, I already did it once with the same notebook, but since yesterday I am receiving an error while importing TRL. I already restarted the kernel and did factory reset, but the error still happens.

from transformers import AutoModelForCausalLM, AutoTokenizer, BitsAndBytesConfig,HfArgumentParser,TrainingArguments,pipeline, logging
from peft import LoraConfig, PeftModel, prepare_model_for_kbit_training, get_peft_model
import os,torch, wandb
from datasets import load_dataset
from trl import SFTTrainer

This is the error:
File /opt/conda/lib/python3.10/site-packages/trl/trainer/init.py:44
42 from .ppo_trainer import PPOTrainer
43 from .reward_trainer import RewardTrainer, compute_accuracy
ā€”> 44 from .sft_trainer import SFTTrainer
45 from .training_configs import RewardConfig

File /opt/conda/lib/python3.10/site-packages/trl/trainer/sft_trainer.py:23
21 import torch.nn as nn
22 from datasets import Dataset
ā€”> 23 from datasets.arrow_writer import SchemaInferenceError
24 from datasets.builder import DatasetGenerationError
25 from transformers import (
26 AutoModelForCausalLM,
27 AutoTokenizer,
(ā€¦)
33 TrainingArguments,
34 )

ImportError: cannot import name ā€˜SchemaInferenceErrorā€™ from ā€˜datasets.arrow_writerā€™ (/opt/conda/lib/python3.10/site-packages/datasets/arrow_writer.py)

2

+1 have the sane issue in Kaggle for Mistral-7B fine tuning

3

@DrPalmiere I fixed it in Kaggle notebook installing latest version of datasets package. Seems, that by default older version of datasets package is used without SchemaInferenceError definition. Use following

%pip install datasets==2.16.0

2 Likes
4

Thanks, I am using this notebook too (Mistral-7B fine tuning).

I tried to update the datasets library but I got the same error. I fix the problem by creating a fresh notebook and copying all the code from the old notebook to the new one, and worked.

2 Likes
5

I tried that but Iā€™m still getting the same error:
ImportError: cannot import name ā€˜SchemaInferenceErrorā€™ from ā€˜datasets.arrow_writerā€™ (/opt/conda/lib/python3.10/site-packages/datasets/arrow_writer.py)

6

Did you try to create an empty notebook and copy cell by cell from the seed notebook? I did it and it is working now?

7

I have done that, I also upgraded the datasets to 2.16.1, But the problem remained.

8

You guys should try this, it works for me

The tip is from:

"In kaggle I used:

  • %pip install -U datasets
    and then restarted runtime and then everything works fine."
1 Like
9

I tried it, but didnā€™ t work. I only fixed it by starting a fresh notebook and doing copy and paste from old notebook to the new one.

1 Like
10

This worked for me :slight_smile:

11

!pip install datasets==2.16.0 -q worked for me