lemone-embed-m

<img src="assets/thumbnail.webp">

Lemone-Embed: A Series of Fine-Tuned Embedding Models for French Taxation

<div class="not-prose bg-gradient-to-r from-gray-50-to-white text-gray-900 border" style="border-radius: 8px; padding: 0.5rem 1rem;"> <p>This series is made up of 7 models, 3 basic models of different sizes trained on 1 epoch, 3 models trained on 2 epochs making up the Boost series and a Pro model with a non-Roberta architecture.</p> </div>

This sentence transformers model, specifically designed for French taxation, has been fine-tuned on a dataset comprising 43 million tokens, integrating a blend of semi-synthetic and fully synthetic data generated by GPT-4 Turbo and Llama 3.1 70B, which have been further refined through evol-instruction tuning and manual curation.

The model is tailored to meet the specific demands of information retrieval across large-scale tax-related corpora, supporting the implementation of production-ready Retrieval-Augmented Generation (RAG) applications. Its primary purpose is to enhance the efficiency and accuracy of legal processes in the taxation domain, with an emphasis on delivering consistent performance in real-world settings, while also contributing to advancements in legal natural language processing research.

This is a sentence-transformers model finetuned from intfloat/multilingual-e5-base. It maps sentences & paragraphs to a 768-dimensional dense vector space and can be used for semantic textual similarity, semantic search, paraphrase mining, text classification, clustering, and more.

Model Details

Model Description

• Model Type: Sentence Transformer
• Base model: intfloat/multilingual-e5-base <!-- at revision d13f1b27baf31030b7fd040960d60d909913633f -->
• Maximum Sequence Length: 512 tokens
• Output Dimensionality: 768 tokens
• Similarity Function: Cosine Similarity
• Developed by: Louis Brulé Naudet
• Funded by: Microsoft for Startups
• Shared by: Louis Brulé Naudet
• Model type: Sentence Transformers
• Language(s) (NLP): FR
• License: Apache 2
• Finetuned from model: intfloat/multilingual-e5-base

Model Sources

• Documentation: Sentence Transformers Documentation
• Repository: Sentence Transformers on GitHub
• Hugging Face: Sentence Transformers on Hugging Face

Full Model Architecture

SentenceTransformer(
  (0): Transformer({'max_seq_length': 512, 'do_lower_case': False}) with Transformer model: XLMRobertaModel 
  (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': False, 'pooling_mode_mean_tokens': True, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
  (2): Normalize()
)

Usage

Direct Usage (Sentence Transformers)

First install the Sentence Transformers library:

pip install -U sentence-transformers

Then you can load this model and run inference.

from sentence_transformers import SentenceTransformer

# Download from the 🤗 Hub
model = SentenceTransformer("louisbrulenaudet/lemone-embed-m")
# Run inference
sentences = [
    "query: Décrivez avec précision les étapes détaillées requises pour traiter les réclamations collectives résultant de désastres agricoles comme définies dans l'article R*198-2 du Livre des procédures fiscales.",
    "passage: Selon l'article R*198-2 du Livre des procédures fiscales, le traitement des réclamations collectives en cas de catastrophes naturelles impactant les cultures agricoles, incluant des phénomènes tels que la grêle ou les inondations, exige la collaboration de plusieurs entités administratives. Initialement, deux commissaires sont nommés par l'administration fiscale pour superviser le processus. Ils sont assistés par un délégué de l'administration des impôts. Avant toute action, le maire de la commune affectée est notifié au moins dix jours avant l'inspection prévue, et il est chargé de communiquer cette date aux résidents via des affichages publics. Les agriculteurs affectés doivent alors rapporter leurs pertes à la mairie avant la réalisation d'un constat officiel par l'inspecteur des impôts, qui sera consigné dans un procès-verbal. Une fois ce document clôturé, aucune réclamation supplémentaire n'est acceptée.",
    "passage: Selon les dispositions de l'article R*196-6 du Livre des procédures fiscales, il est attribué aux sujets fiscaux un intervalle précisément défini pour élever des réclamations à l'égard des taxes, cotisations et autres prélèvements relatifs aux céréales et leurs transformés. Ce délai se prolonge jusqu'à la fin de la campagne agricole suivante celle au cours de laquelle l'avis de mise en recouvrement de la taxe a été notifié ou le règlement de l'imposition contestée effectué, permettant ainsi aux parties prenantes de se prévaloir de leurs prérogatives contestataires avec une certitude temporelle.",
]
embeddings = model.encode(sentences)
print(embeddings.shape)
# [3, 768]

# Get the similarity scores for the embeddings
similarities = model.similarity(embeddings, embeddings)
print(similarities.shape)
# [3, 3]

<!-- ### Direct Usage (Transformers) <details><summary>Click to see the direct usage in Transformers</summary> </details> --> <!-- ### Downstream Usage (Sentence Transformers) You can finetune this model on your own dataset. <details><summary>Click to expand</summary> </details> --> <!-- ### Out-of-Scope Use *List how the model may foreseeably be misused and address what users ought not to do with the model.* -->

Evaluation

Metrics

Information Retrieval

• Dataset: Lemone-information-retrieval
• Evaluated with <code>InformationRetrievalEvaluator</code>

<!-- ## Bias, Risks and Limitations *What are the known or foreseeable issues stemming from this model? You could also flag here known failure cases or weaknesses of the model.* --> <!-- ### Recommendations *What are recommendations with respect to the foreseeable issues? For example, filtering explicit content.* -->

Training Details

Training Dataset

• Size: 296,234 training samples
• Columns: <code>query</code>, <code>positive</code>, and <code>negative</code>
• Approximate statistics based on the first 1000 samples:

  	query	positive	negative
  type	string	string	string
  details	<ul><li>min: 24 tokens</li><li>mean: 54.2 tokens</li><li>max: 179 tokens</li></ul>	<ul><li>min: 75 tokens</li><li>mean: 182.28 tokens</li><li>max: 332 tokens</li></ul>	<ul><li>min: 53 tokens</li><li>mean: 190.2 tokens</li><li>max: 456 tokens</li></ul>

• Loss: <code>CachedGISTEmbedLoss</code> with these parameters:

  {'guide': SentenceTransformer(
    (0): Transformer({'max_seq_length': 8192, 'do_lower_case': False}) with Transformer model: NewModel 
    (1): Pooling({'word_embedding_dimension': 768, 'pooling_mode_cls_token': True, 'pooling_mode_mean_tokens': False, 'pooling_mode_max_tokens': False, 'pooling_mode_mean_sqrt_len_tokens': False, 'pooling_mode_weightedmean_tokens': False, 'pooling_mode_lasttoken': False, 'include_prompt': True})
    (2): Normalize()
  ), 'temperature': 0.01}
  

Training Hyperparameters

Non-Default Hyperparameters

• eval_strategy: steps
• per_device_train_batch_size: 512
• learning_rate: 2e-05
• num_train_epochs: 1
• warmup_ratio: 0.1
• fp16: True
• batch_sampler: no_duplicates

All Hyperparameters

<details><summary>Click to expand</summary>

• overwrite_output_dir: False
• do_predict: False
• eval_strategy: steps
• prediction_loss_only: True
• per_device_train_batch_size: 512
• per_device_eval_batch_size: 8
• per_gpu_train_batch_size: None
• per_gpu_eval_batch_size: None
• gradient_accumulation_steps: 1
• eval_accumulation_steps: None
• torch_empty_cache_steps: None
• learning_rate: 2e-05
• weight_decay: 0.0
• adam_beta1: 0.9
• adam_beta2: 0.999
• adam_epsilon: 1e-08
• max_grad_norm: 1.0
• num_train_epochs: 1
• max_steps: -1
• lr_scheduler_type: linear
• lr_scheduler_kwargs: {}
• warmup_ratio: 0.1
• warmup_steps: 0
• log_level: passive
• log_level_replica: warning
• log_on_each_node: True
• logging_nan_inf_filter: True
• save_safetensors: True
• save_on_each_node: False
• save_only_model: False
• restore_callback_states_from_checkpoint: False
• no_cuda: False
• use_cpu: False
• use_mps_device: False
• seed: 42
• data_seed: None
• jit_mode_eval: False
• use_ipex: False
• bf16: False
• fp16: True
• fp16_opt_level: O1
• half_precision_backend: auto
• bf16_full_eval: False
• fp16_full_eval: False
• tf32: None
• local_rank: 0
• ddp_backend: None
• tpu_num_cores: None
• tpu_metrics_debug: False
• debug: []
• dataloader_drop_last: False
• dataloader_num_workers: 0
• dataloader_prefetch_factor: None
• past_index: -1
• disable_tqdm: False
• remove_unused_columns: True
• label_names: None
• load_best_model_at_end: False
• ignore_data_skip: False
• fsdp: []
• fsdp_min_num_params: 0
• fsdp_config: {'min_num_params': 0, 'xla': False, 'xla_fsdp_v2': False, 'xla_fsdp_grad_ckpt': False}
• fsdp_transformer_layer_cls_to_wrap: None
• accelerator_config: {'split_batches': False, 'dispatch_batches': None, 'even_batches': True, 'use_seedable_sampler': True, 'non_blocking': False, 'gradient_accumulation_kwargs': None}
• deepspeed: None
• label_smoothing_factor: 0.0
• optim: adamw_torch
• optim_args: None
• adafactor: False
• group_by_length: False
• length_column_name: length
• ddp_find_unused_parameters: None
• ddp_bucket_cap_mb: None
• ddp_broadcast_buffers: False
• dataloader_pin_memory: True
• dataloader_persistent_workers: False
• skip_memory_metrics: True
• use_legacy_prediction_loop: False
• push_to_hub: False
• resume_from_checkpoint: None
• hub_model_id: None
• hub_strategy: every_save
• hub_private_repo: False
• hub_always_push: False
• gradient_checkpointing: False
• gradient_checkpointing_kwargs: None
• include_inputs_for_metrics: False
• eval_do_concat_batches: True
• fp16_backend: auto
• push_to_hub_model_id: None
• push_to_hub_organization: None
• mp_parameters:
• auto_find_batch_size: False
• full_determinism: False
• torchdynamo: None
• ray_scope: last
• ddp_timeout: 1800
• torch_compile: False
• torch_compile_backend: None
• torch_compile_mode: None
• dispatch_batches: None
• split_batches: None
• include_tokens_per_second: False
• include_num_input_tokens_seen: False
• neftune_noise_alpha: None
• optim_target_modules: None
• batch_eval_metrics: False
• eval_on_start: False
• eval_use_gather_object: False
• batch_sampler: no_duplicates
• multi_dataset_batch_sampler: proportional

</details>

Environmental Impact

Carbon emissions were measured using CodeCarbon.

• Energy Consumed: 1.243 kWh
• Carbon Emitted: 0.459 kg of CO2
• Hours Used: 2.244 hours

Training Hardware

• On Cloud: No
• GPU Model: 1 x NVIDIA H100 NVL
• CPU Model: AMD EPYC 9V84 96-Core Processor
• RAM Size: 314.69 GB

Framework Versions

• Python: 3.10.12
• Sentence Transformers: 3.1.1
• Transformers: 4.44.2
• PyTorch: 2.3.0+cu121
• Accelerate: 0.33.0
• Datasets: 2.21.0
• Tokenizers: 0.19.1

Citation

BibTeX

Sentence Transformers

@inproceedings{reimers-2019-sentence-bert,
    title = "Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks",
    author = "Reimers, Nils and Gurevych, Iryna",
    booktitle = "Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing",
    month = "11",
    year = "2019",
    publisher = "Association for Computational Linguistics",
    url = "https://arxiv.org/abs/1908.10084",
}

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

@misc{louisbrulenaudet2024,
  author =       {Louis Brulé Naudet},
  title =        {Lemone-Embed: A Series of Fine-Tuned Embedding Models for French Taxation},
  year =         {2024}
  howpublished = {\url{https://huggingface.co/datasets/louisbrulenaudet/lemone-embed-m}},
}

Feedback

If you have any feedback, please reach out at [email protected].