Skip to content

Latest commit

 

History

History
108 lines (72 loc) · 4.15 KB

README.md

File metadata and controls

108 lines (72 loc) · 4.15 KB
Raw

Location Mention Recognition (LMR) Fine-Tuning Script

Overview

This repository contains a script designed for fine-tuning the GLiNER model to recognize and classify location mentions within text data. The script utilizes the Hugging Face GLiNER library and includes customizations for data processing, model training, and evaluation. It supports training with large, pre-trained GLiNER models, which have been specifically tailored for Named Entity Recognition (NER) tasks like location recognition.

Requirements

  • Python 3.8+
  • CUDA-enabled GPU for model training
  • Necessary Python packages (install via requirements.txt):
    • transformers
    • torch
    • datasets
    • wandb
    • gliner

Environment Variables

This script requires several environment variables for managing resources and logging during model training:

  • TOKENIZERS_PARALLELISM: Controls parallelism in tokenization to improve performance.
  • CUDA_DEVICE_ORDER: Ensures that CUDA devices are initialized in a consistent order.
  • CUDA_VISIBLE_DEVICES: Specifies which GPU to use for training.
  • WANDB_PROJECT: Defines the project name for Weights and Biases logging.
  • WANDB_WATCH & WANDB_NOTEBOOK_NAME: Used for logging with Weights and Biases.

Dataset

The training and test datasets should be JSON files containing pre-processed location mention data. The script expects the following format:

[
    {
        "tokens": [...],
        "ner": [[start_index, end_index, "LOC"]],
        "label": ["location"]
    }
]

Dataset Paths:

  • Training Data: data/accepted_data/TrainCleaned.json
  • Test Data: data/accepted_data/TestCleaned.json

The script automatically updates the NER labels to "location".

Zindi dataset to train this model can be found here

Usage

Model Setup

The script can load several pre-trained GLiNER models for fine-tuning, including:

  • urchade/gliner_large-v2.1 (default)
  • urchade/gliner_small-v2.1
  • urchade/gliner_medium-v2.1

You can select a different model by changing the GLiNER.from_pretrained() method in the code.

Fine-Tuning

The script fine-tunes the selected model on the provided dataset, using the Trainer class from the GLiNER library. Several training parameters, such as batch size, learning rate, and number of epochs, can be configured. Key parameters include:

  • Batch Size: Set to 8 by default.
  • Learning Rate: Default is 1e-6 with weight decay for regularization.
  • Epochs: Number of epochs for training, set to 5.
  • Evaluation Strategy: Evaluation is done periodically based on the number of steps.
  • Save Strategy: Models are saved every save_steps interval, with the best model loaded at the end.

Execution

To start training, run:

python <script_name>.py

Logging

The script uses Weights and Biases for logging training progress, loss metrics, and saving checkpoints.

To use Weights and Biases, ensure you have set up a project by setting the environment variable WANDB_PROJECT and logging in via wandb login.

Customization

  • Data Collator: The data collator is configured to prepare labels for location mentions. You can modify this if you need to handle different types of entities.
  • Trainer Arguments: You can fine-tune other hyperparameters in the TrainingArguments section such as gradient accumulation steps, learning rate scheduler, and evaluation strategy.
  • Model: Various GLiNER models are available for specific use cases. To try a different pre-trained model, change the model's path in the from_pretrained() method.

Example Output

The script will output:

  1. Model summary.
  2. Dataset size information.
  3. Sample dataset entries.
  4. Training logs with loss metrics.
  5. Model evaluation results.

Notes

  • This script is optimized for large datasets and uses a constant learning rate for fine-tuning. Adjust the learning rate or use a different scheduler (e.g., cosine, linear) depending on your dataset and hardware.
  • The script is configured to load the best model at the end of training based on evaluation loss.