bofip / README.md
louisbrulenaudet's picture
Update README.md
399719d
|
raw
history blame
No virus
5.79 kB
metadata
license: apache-2.0
language:
  - fr
multilinguality:
  - monolingual
tags:
  - finetuning
  - legal
  - french law
  - droit français
  - Bofip
source_datasets:
  - original
pretty_name: Bulletin officiel des finances publiques - impôts
task_categories:
  - text-generation
  - table-question-answering
  - summarization
  - conversational
size_categories:
  - 1K<n<10K

Bulletin officiel des finances publiques - impôts, non-instruct (11-12-2023)

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.

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.

We used the following list of instructions for generating the dataset:

instructions = [
    "Compose l'intégralité de la publication sous forme écrite.",
    "Écris la totalité du contenu de la publication.",
    "Formule la totalité du texte présent dans la publication.",
    "Produis l'intégralité de la publication en écriture.",
    "Développe la publication dans son ensemble par écrit.",
    "Génère l'ensemble du texte contenu dans la publication.",
    "Formule le contenu intégral de la publication en entier.",
    "Rédige la totalité du texte de la publication en entier.",
    "Compose l'intégralité du contenu textuel de la publication.",
    "Rédige l'ensemble du texte qui constitue la publication.",
    "Composez l'intégralité de la publication sous forme écrite.",
    "Écrivez la totalité du contenu de la publication.",
    "Formulez la totalité du texte présent dans la publication.",
    "Développez la publication dans son ensemble par écrit.",
    "Générez l'ensemble du texte contenu dans la publication.",
    "Formulez le contenu intégral de la publication en entier.",
    "Rédigez la totalité du texte de la publication en entier.",
    "Composez l'intégralité du contenu textuel de la publication.",
    "Écrivez la publication dans son intégralité en termes de texte.",
    "Rédigez l'ensemble du texte qui constitue la publication.",
    "Composer l'intégralité de la publication sous forme écrite.",
    "Écrire la totalité du contenu de la publication.",
    "Formuler la totalité du texte présent dans la publication.",
    "Produire l'intégralité de la publication en écriture.",
    "Développer la publication dans son ensemble par écrit.",
    "Générer l'ensemble du texte contenu dans la publication.",
    "Formuler le contenu intégral de la publication en entier.",
    "Rédiger la totalité du texte de la publication en entier.",
    "Composer l'intégralité du contenu textuel de la publication.",
    "Rédiger l'ensemble du texte qui constitue la publication.",
    "Quelles sont les dispositions de la publication ?",
    "Quelles dispositions sont incluses dans la publication ?",
    "Quelles sont les dispositions énoncées dans la publication ?",
    "Quel est le texte intégral de la publication ?",
    "Quelle est la lettre de la publication ?"
]

Citing this project

If you use this code in your research, please use the following BibTeX entry.

@misc{louisbrulenaudet2023,
  author =       {Louis Brulé Naudet},
  title =        {Bulletin officiel des finances publiques - impôts, non-instruct (11-12-2023)},
  howpublished = {\url{https://huggingface.co/datasets/louisbrulenaudet/bofip}},
  year =         {2023}
}

Feedback

If you have any feedback, please reach out at louisbrulenaudet@icloud.com.