LLaMA Model Fix for Kaggle Platform
Budget: $10 – $30 USD
Project Description:
We are working on a text classification project using the LLaMA model on the Kaggle platform. The objective of the project is to classify a dataset into different categories based on text inputs. The pipeline was previously functioning correctly; however, a recent update in the SFTTrainer method (or related library) has introduced a breaking change.
Current Issue:
Problem:
The issue arises when passing the tokenizer to the SFTTrainer. Due to an update, the current implementation (which attempts to access a tokenizer attribute on the tokenizer instance) now fails with an error:
AttributeError: PreTrainedTokenizerFast has no attribute tokenizer.
Background:
The project was fully functional before this update. The SFTTrainer previously accepted the tokenizer in a certain format, but the update requires that the tokenizer be passed directly (or via an updated parameter such as processing_class), without trying to access a non-existent nested attribute.
Project Goals:
Primary Objective:
Modify the SFTTrainer initialization so that it correctly accepts the tokenizer, allowing the model to train properly and enabling extraction of output for text classification.
Specific Tasks:
Analyze the current implementation and error.
Update the code to pass the tokenizer correctly (e.g., directly pass the tokenizer instance to processing_class).
Ensure the model (LLaMA) fine-tunes properly on the dataset and classifies text into different categories.
Test the solution thoroughly on the Kaggle platform to confirm that the training process completes successfully and outputs are correctly extracted.
Additional Information:
Platform:
The project is built and run on the Kaggle platform.
Model & Dataset:
The project uses the LLaMA model for text classification. The dataset consists of text samples that need to be categorized into different classes.
I hope to being speak Arabic to share the details
We are working on a text classification project using the LLaMA model on the Kaggle platform. The objective of the project is to classify a dataset into different categories based on text inputs. The pipeline was previously functioning correctly; however, a recent update in the SFTTrainer method (or related library) has introduced a breaking change.
Current Issue:
Problem:
The issue arises when passing the tokenizer to the SFTTrainer. Due to an update, the current implementation (which attempts to access a tokenizer attribute on the tokenizer instance) now fails with an error:
AttributeError: PreTrainedTokenizerFast has no attribute tokenizer.
Background:
The project was fully functional before this update. The SFTTrainer previously accepted the tokenizer in a certain format, but the update requires that the tokenizer be passed directly (or via an updated parameter such as processing_class), without trying to access a non-existent nested attribute.
Project Goals:
Primary Objective:
Modify the SFTTrainer initialization so that it correctly accepts the tokenizer, allowing the model to train properly and enabling extraction of output for text classification.
Specific Tasks:
Analyze the current implementation and error.
Update the code to pass the tokenizer correctly (e.g., directly pass the tokenizer instance to processing_class).
Ensure the model (LLaMA) fine-tunes properly on the dataset and classifies text into different categories.
Test the solution thoroughly on the Kaggle platform to confirm that the training process completes successfully and outputs are correctly extracted.
Additional Information:
Platform:
The project is built and run on the Kaggle platform.
Model & Dataset:
The project uses the LLaMA model for text classification. The dataset consists of text samples that need to be categorized into different classes.
I hope to being speak Arabic to share the details