The dataset viewer is not available for this split.
Cannot extract the features (columns) for the split 'train' of the config 'default' of the dataset.
Error code:   FeaturesError
Exception:    ValueError
Message:      Not able to read records in the JSON file at zip://dataset.full.json::hf://datasets/lemoneresearch/cgi@b6e602ed7e1ce9f89d150301ae8fe68b541ba946/dataset.full.json.zip.
Traceback:    Traceback (most recent call last):
                File "/src/services/worker/src/worker/job_runners/split/first_rows.py", line 241, in compute_first_rows_from_streaming_response
                  iterable_dataset = iterable_dataset._resolve_features()
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 2216, in _resolve_features
                  features = _infer_features_from_batch(self.with_format(None)._head())
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 1239, in _head
                  return _examples_to_batch(list(self.take(n)))
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 1389, in __iter__
                  for key, example in ex_iterable:
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 1044, in __iter__
                  yield from islice(self.ex_iterable, self.n)
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/iterable_dataset.py", line 282, in __iter__
                  for key, pa_table in self.generate_tables_fn(**self.kwargs):
                File "/src/services/worker/.venv/lib/python3.9/site-packages/datasets/packaged_modules/json/json.py", line 165, in _generate_tables
                  raise ValueError(f"Not able to read records in the JSON file at {file}.") from None
              ValueError: Not able to read records in the JSON file at zip://dataset.full.json::hf://datasets/lemoneresearch/cgi@b6e602ed7e1ce9f89d150301ae8fe68b541ba946/dataset.full.json.zip.

Need help to make the dataset viewer work? Open a discussion for direct support.

YAML Metadata Warning: The task_categories "conversational" is not in the official list: text-classification, token-classification, table-question-answering, question-answering, zero-shot-classification, translation, summarization, feature-extraction, text-generation, text2text-generation, fill-mask, sentence-similarity, text-to-speech, text-to-audio, automatic-speech-recognition, audio-to-audio, audio-classification, voice-activity-detection, depth-estimation, image-classification, object-detection, image-segmentation, text-to-image, image-to-text, image-to-image, image-to-video, unconditional-image-generation, video-classification, reinforcement-learning, robotics, tabular-classification, tabular-regression, tabular-to-text, table-to-text, multiple-choice, text-retrieval, time-series-forecasting, text-to-video, image-text-to-text, visual-question-answering, document-question-answering, zero-shot-image-classification, graph-ml, mask-generation, zero-shot-object-detection, text-to-3d, image-to-3d, image-feature-extraction, other

Code Général des Impôts, non-instruct (11-12-2023)

This project focuses on fine-tuning pre-trained language models to create efficient and accurate models for tax 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:

  • version: string, denoting the version associated with the element.
  • 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.
  • complexity: int, reflecting the degree of abstraction requested from the LLM (Legal Language Model). A value of 1 represents an instruction grounded in authoritative text, while 2 introduces added complexity or abstraction.
  • created_at: date, capturing the date and time of the document's creation.
  • updated_at: date, detailing the most recent update's date and time.
  • expiration: date, delineating the expiration date of the legal information.
  • status: string, specifying the application status of the law.
  • coming_into_force: date, signifying the date when the legal information becomes enforceable.
  • language: string, describing the language in which the legal information is presented.
  • length: int, offering information regarding the length of the legal content.
  • source: string, representing the source from which the legal information originated.

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

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 ?"
]

Citing this project

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

@misc{louisbrulenaudet2023,
  author =       {Louis Brulé Naudet},
  title =        {Code Général des Impôts, non-instruct (11-12-2023)},
  howpublished = {\url{https://huggingface.co/datasets/louisbrulenaudet/cgi}},
  year =         {2023}
}

Feedback

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

Downloads last month
0
Edit dataset card