---
license: apache-2.0
task_categories:
- summarization
- text-generation
- feature-extraction
tags:
- legal
- droit
- jurisprudence
- cassation
- judilibre
pretty_name: >-
  Training dataset for summarizing and titling decisions of the French Court of
  Cassation
size_categories:
- 10K<n<100K
language:
- fr
---

## Dataset Description
- **Repository:** https://huggingface.co/datasets/maurya/alpaca_ccass_motivations_sommaires_titres
- **Leaderboard:** N/A
- **Point of Contact:** [Amaury Fouret](mailto:amaury.fouret@justice.fr)
  

# Training dataset for summarizing and titling decisions of the French Court of cassation based on motivations

This alpaca-format dataset is designed to train models for summarizing and titling French Supreme Court decisions based on the grounds of them. Created with a view to producing metadata for decisions not published in the bulletin, this dataset aims to simplify the development of annotation and categorization tools, and is positioned as a facilitator for jurisprudential processing and information organization.

## Dataset Structure

The dataset includes hierarchical keyword extraction task that generates structured legal taxonomies from court decisions. The hierarchical keywords follow a specific format using special tokens to denote a hierarchical relationship between keywords of different levels.

## Special Tokens

When fine-tuning models on this dataset, you will need to add the following special tokens to your tokenizer to properly handle the hierarchical keyword structure:

### Required Special Tokens
```python
additional_special_tokens = [
    "<start_hierarchy>", "<end_hierarchy>",
    "<start_level1>", "<end_level1>",
    "<start_level2>", "<end_level2>",
    "<start_level3>", "<end_level3>",
    "<start_level4>", "<end_level4>",
    "<start_level5>", "<end_level5>",
    "<start_level6>", "<end_level6>",
    "<start_level7>", "<end_level7>",
    "<start_level8>", "<end_level8>",
    "<start_level9>", "<end_level9>",
    "<start_level10>", "<end_level10>",
    "<start_level11>", "<end_level11>",
    "<start_level12>", "<end_level12>",
    "<start_level13>", "<end_level13>"
]
```

### Example Output Format
The hierarchical keywords are structured as follows:
```
<start_hierarchy><start_level1>REPRESENTATION DES SALARIES<end_level1><start_level2>Comité d'entreprise<end_level2><start_level3>Fonctionnement<end_level3><start_level4>Réunion<end_level4><start_level5>Présidence<end_level5><end_hierarchy>
```

## Implementation

### Adding Special Tokens to Your Tokenizer
```python
from transformers import AutoTokenizer

# Load your base tokenizer
tokenizer = AutoTokenizer.from_pretrained("your-base-model")

# Add the special tokens
tokenizer.add_tokens(additional_special_tokens, special_tokens=True)

# Don't forget to resize your model's token embeddings
model.resize_token_embeddings(len(tokenizer))
```

### Configuration File Setup
For YAML configuration files:
```yaml
special_tokens:
  bos_token: "<s>"
  eos_token: "</s>"
  unk_token: "<unk>"
  pad_token: "[PAD]"

additional_special_tokens:
  - "<start_hierarchy>"
  - "<end_hierarchy>"
  - "<start_level1>"
  - "<end_level1>"
  # ... (include all tokens up to level 13)
```

## Use Cases

- **Legal Text Summarization**: Generate concise summaries of court decisions
- **Hierarchical Classification**: Extract structured legal taxonomies
- **Metadata Generation**: Create standardized metadata for unpublished decisions
- **Legal Research Tools**: Facilitate automated categorization and search

## Dataset Features

- Alpaca instruction format for easy integration with popular training frameworks
- Hierarchical keyword structures with up to 14 levels of legal classification
- Specialized for French Court of Cassation decisions and legal reasoning patterns

## Feedback

If you have any feedback, please reach out at [Amaury Fouret](mailto:amaury.fouret@justice.fr).