|
--- |
|
license: apache-2.0 |
|
language: |
|
- fr |
|
multilinguality: |
|
- monolingual |
|
tags: |
|
- finetuning |
|
- legal |
|
- french law |
|
- droit français |
|
- Code du patrimoine |
|
source_datasets: |
|
- original |
|
pretty_name: Code du patrimoine |
|
task_categories: |
|
- text-generation |
|
- table-question-answering |
|
- summarization |
|
- text-retrieval |
|
- question-answering |
|
- text-classification |
|
size_categories: |
|
- 1K<n<10K |
|
--- |
|
# Code du patrimoine, non-instruct (2024-06-18) |
|
|
|
This project focuses on fine-tuning pre-trained language models to create efficient and accurate models for legal practice. |
|
|
|
Fine-tuning is the process of adapting a pre-trained model to perform specific tasks or cater to particular domains. It involves adjusting the model's parameters through a further round of training on task-specific or domain-specific data. While conventional fine-tuning strategies involve supervised learning with labeled data, instruction-based fine-tuning introduces a more structured and interpretable approach. |
|
|
|
Instruction-based fine-tuning leverages the power of human-provided instructions to guide the model's behavior. These instructions can be in the form of text prompts, prompts with explicit task descriptions, or a combination of both. This approach allows for a more controlled and context-aware interaction with the LLM, making it adaptable to a multitude of specialized tasks. |
|
|
|
Instruction-based fine-tuning significantly enhances the performance of LLMs in the following ways: |
|
|
|
- Task-Specific Adaptation: LLMs, when fine-tuned with specific instructions, exhibit remarkable adaptability to diverse tasks. They can switch seamlessly between translation, summarization, and question-answering, guided by the provided instructions. |
|
- Reduced Ambiguity: Traditional LLMs might generate ambiguous or contextually inappropriate responses. Instruction-based fine-tuning allows for a clearer and more context-aware generation, reducing the likelihood of nonsensical outputs. |
|
- Efficient Knowledge Transfer: Instructions can encapsulate domain-specific knowledge, enabling LLMs to benefit from expert guidance. This knowledge transfer is particularly valuable in fields like tax practice, law, medicine, and more. |
|
- Interpretability: Instruction-based fine-tuning also makes LLM behavior more interpretable. Since the instructions are human-readable, it becomes easier to understand and control model outputs. |
|
- Adaptive Behavior: LLMs, post instruction-based fine-tuning, exhibit adaptive behavior that is responsive to both explicit task descriptions and implicit cues within the provided text. |
|
|
|
## Concurrent reading of the LegalKit |
|
|
|
To use all the legal data published on LegalKit, you can use this code snippet: |
|
```python |
|
# -*- coding: utf-8 -*- |
|
import concurrent.futures |
|
import os |
|
|
|
import datasets |
|
from tqdm.notebook import tqdm |
|
|
|
def dataset_loader( |
|
name:str, |
|
streaming:bool=True |
|
) -> datasets.Dataset: |
|
""" |
|
Helper function to load a single dataset in parallel. |
|
|
|
Parameters |
|
---------- |
|
name : str |
|
Name of the dataset to be loaded. |
|
|
|
streaming : bool, optional |
|
Determines if datasets are streamed. Default is True. |
|
|
|
Returns |
|
------- |
|
dataset : datasets.Dataset |
|
Loaded dataset object. |
|
|
|
Raises |
|
------ |
|
Exception |
|
If an error occurs during dataset loading. |
|
""" |
|
try: |
|
return datasets.load_dataset( |
|
name, |
|
split="train", |
|
streaming=streaming |
|
) |
|
|
|
except Exception as exc: |
|
logging.error(f"Error loading dataset {name}: {exc}") |
|
|
|
return None |
|
|
|
|
|
def load_datasets( |
|
req:list, |
|
streaming:bool=True |
|
) -> list: |
|
""" |
|
Downloads datasets specified in a list and creates a list of loaded datasets. |
|
|
|
Parameters |
|
---------- |
|
req : list |
|
A list containing the names of datasets to be downloaded. |
|
|
|
streaming : bool, optional |
|
Determines if datasets are streamed. Default is True. |
|
|
|
Returns |
|
------- |
|
datasets_list : list |
|
A list containing loaded datasets as per the requested names provided in 'req'. |
|
|
|
Raises |
|
------ |
|
Exception |
|
If an error occurs during dataset loading or processing. |
|
|
|
Examples |
|
-------- |
|
>>> datasets = load_datasets(["dataset1", "dataset2"], streaming=False) |
|
""" |
|
datasets_list = [] |
|
|
|
with concurrent.futures.ThreadPoolExecutor() as executor: |
|
future_to_dataset = {executor.submit(dataset_loader, name): name for name in req} |
|
|
|
for future in tqdm(concurrent.futures.as_completed(future_to_dataset), total=len(req)): |
|
name = future_to_dataset[future] |
|
|
|
try: |
|
dataset = future.result() |
|
|
|
if dataset: |
|
datasets_list.append(dataset) |
|
|
|
except Exception as exc: |
|
logging.error(f"Error processing dataset {name}: {exc}") |
|
|
|
return datasets_list |
|
|
|
|
|
req = [ |
|
"louisbrulenaudet/code-artisanat", |
|
"louisbrulenaudet/code-action-sociale-familles", |
|
# ... |
|
] |
|
|
|
datasets_list = load_datasets( |
|
req=req, |
|
streaming=True |
|
) |
|
|
|
dataset = datasets.concatenate_datasets( |
|
datasets_list |
|
) |
|
``` |
|
|
|
## Dataset generation |
|
|
|
This JSON file is a list of dictionaries, each dictionary contains the following fields: |
|
|
|
- `instruction`: `string`, presenting the instruction linked to the element. |
|
- `input`: `string`, signifying the input details for the element. |
|
- `output`: `string`, indicating the output information for the element. |
|
- `start`: `string`, the date of entry into force of the article. |
|
- `expiration`: `string`, the date of expiration of the article. |
|
- `num`: `string`, the id of the article. |
|
|
|
We used the following list of instructions for generating the dataset: |
|
```python |
|
instructions = [ |
|
"Compose l'intégralité de l'article sous forme écrite.", |
|
"Écris la totalité du contenu de l'article.", |
|
"Formule la totalité du texte présent dans l'article.", |
|
"Produis l'intégralité de l'article en écriture.", |
|
"Développe l'article dans son ensemble par écrit.", |
|
"Génère l'ensemble du texte contenu dans l'article.", |
|
"Formule le contenu intégral de l'article en entier.", |
|
"Rédige la totalité du texte de l'article en entier.", |
|
"Compose l'intégralité du contenu textuel de l'article.", |
|
"Rédige l'ensemble du texte qui constitue l'article.", |
|
"Formule l'article entier dans son contenu écrit.", |
|
"Composez l'intégralité de l'article sous forme écrite.", |
|
"Écrivez la totalité du contenu de l'article.", |
|
"Formulez la totalité du texte présent dans l'article.", |
|
"Développez l'article dans son ensemble par écrit.", |
|
"Générez l'ensemble du texte contenu dans l'article.", |
|
"Formulez le contenu intégral de l'article en entier.", |
|
"Rédigez la totalité du texte de l'article en entier.", |
|
"Composez l'intégralité du contenu textuel de l'article.", |
|
"Écrivez l'article dans son intégralité en termes de texte.", |
|
"Rédigez l'ensemble du texte qui constitue l'article.", |
|
"Formulez l'article entier dans son contenu écrit.", |
|
"Composer l'intégralité de l'article sous forme écrite.", |
|
"Écrire la totalité du contenu de l'article.", |
|
"Formuler la totalité du texte présent dans l'article.", |
|
"Produire l'intégralité de l'article en écriture.", |
|
"Développer l'article dans son ensemble par écrit.", |
|
"Générer l'ensemble du texte contenu dans l'article.", |
|
"Formuler le contenu intégral de l'article en entier.", |
|
"Rédiger la totalité du texte de l'article en entier.", |
|
"Composer l'intégralité du contenu textuel de l'article.", |
|
"Rédiger l'ensemble du texte qui constitue l'article.", |
|
"Formuler l'article entier dans son contenu écrit.", |
|
"Quelles sont les dispositions de l'article ?", |
|
"Quelles dispositions sont incluses dans l'article ?", |
|
"Quelles sont les dispositions énoncées dans l'article ?", |
|
"Quel est le texte intégral de l'article ?", |
|
"Quelle est la lettre de l'article ?" |
|
] |
|
``` |
|
|
|
## Feedback |
|
|
|
If you have any feedback, please reach out at [louisbrulenaudet@icloud.com](mailto:louisbrulenaudet@icloud.com). |
