llm-extractinator 0.3.2

A framework that enables efficient extraction of structured data from unstructured text using large language models (LLMs).

LLM Extractinator

[!Important] This tool is a prototype which is in active development and is still undergoing major changes. Please always check the results!

---

Overview

This project enables the efficient extraction of structured data from unstructured text using large language models (LLMs). It provides a flexible configuration system and supports a variety of tasks.

Tool Workflow

---

Installing Ollama

For the package to work, Ollama needs to be installed on your machine. For Linux, use the following command:

curl -fsSL https://ollama.com/install.sh | sh

For Windows or macOS, install via this link

Installing the Package

Option 1: Install from PyPI

The package is installable via PyPI using:

pip install llm_extractinator

Option 2: Install using local clone

For contributing to or developing the package, clone this repository and install it using:

pip install -e .

---

Data Structure

The data structure for the input data should be as follows:

• The data should be in a CSV or a JSON file.
• The text data should be in a column specified by the Input_Field in the task configuration. The default is text.
• The name of the data file should be specified in the Data_Path field in the task configuration. The default location is the data folder, but this can be changed using the --data_dir flag when running the model.

When running the model with examples (num_examples > 0), the examples should be provided in a separate CSV or JSON file. The path to this file should be specified in the Example_Path field in the task configuration. The default location is the data folder, but this can be changed using the --example_dir flag when running the model.

---

Setting Up Task Configuration

Create a JSON file in the tasks folder for each task, following the naming convention:

TaskXXX_taskname.json

Where XXX is a 3-digit number, and taskname is a brief descriptor of the task.

The JSON file should always include the following fields:

• Task: The name of the task.
• Type: The type of task.
• Description: A detailed description of the task.
• Data_Path: The filename of the data file in the data folder.
• Input_Field: The column name containing the text data.
• Parser_Format: The JSON format you want the output to be in. See Task999_example.json for an example.

The following field is only mandatory if you want to have the model use examples in its prompt:

• Example_Path: The path to data used for creating examples (only required if num_examples > 0 when running the model).

[!Important] If you don't want to use examples, omit the Example_Path field from the task configuration completely. Do not set it to an empty string!

---

Input Flags for extractinate

The following input flags can be used to configure the behavior of the extractinate script:

Flag	Type	Default Value	Description
--task_id	int	Required	Task ID to generate examples for.
--run_name	str	"run"	Name of the run for logging purposes.
--n_runs	int	5	Number of runs to perform.
--num_examples	int	0	Number of examples to generate for each task.
--num_predict	int	512	Maximum number of tokens to predict.
--chunk_size	int	None	Number of examples to generate in a single chunk. When None, use dataset size as chunksize.
--overwrite	bool	False	Overwrite existing files instead of skipping them.
--translate	bool	False	Translate the generated examples to English.
--verbose	bool	False	Enable verbose logging.
--reasoning_model	bool	False	Whether or not the model is a reasoning model.
--model_name	str	"mistral-nemo"	Name of the model to use for prediction tasks.
--temperature	float	0.0	Temperature for text generation.
--max_context_len	str	max	Maximum context length; 'split' splits data into short and long cases and does a run for them seperately (good if your dataset distribution has a tail with long reports and a bulk of short ones), 'max' uses the maximum token length of the dataset, or a number sets a fixed length.
--top_k	int	None	Limits the sampling to the top K tokens.
--top_p	float	None	Nucleus sampling probability threshold.
--seed	int	None	Random seed for reproducibility.
--output_dir	Path	<project_root>/output	Path to the directory for output files.
--task_dir	Path	<project_root>/tasks	Path to the directory containing task configuration files.
--log_dir	Path	<project_root>/output	Path to the directory for log files.
--data_dir	Path	<project_root>/data	Path to the directory containing input data.
--example_dir	Path	<project_root>/examples	Path to the directory containing example data.
--host	str	"localhost"	Server hostname.
--port	int	28900	Server port.

---

Running the Extractor

To run the data extraction process, use either the command line or import the function in Python.

Using the Command Line

extractinate --task_id 001 --model_name "mistral-nemo" 

Using the Function in Python

from llm_extractinator import extractinate

extractinate(
    task_id=1,
    model_name="mistral-nemo",
)

---

Enhancements and Contributions

Feel free to contribute by improving configurations, adding more task types, or extending model compatibility. Open a pull request or file an issue for discussions!