inclusively-classification

Technical Overview

The inclusively‑classification model (model ID E-MIMIC/inclusively-classification) is a fine‑tuned Italian BERT classifier that detects the presence of inclusive language in Italian sentences. Built on top of the large‑scale Italian BERT (dbmdz/bert-base-italian-xxl-cased), it maps each input text to one of three mutually exclusive categories:

• inclusive – the sentence uses inclusive wording (e.g., “Il personale docente e non docente”).
• not_inclusive – the sentence is non‑inclusive (e.g., “I professori”).
• not_pertinent – the sentence does not pertain to the task (e.g., “La scuola è chiusa”).

Key features include:

• Native Italian language support with a cased vocabulary of 30 k tokens.
• Optimized for short‑to‑medium length inputs (max length = 128 tokens).
• Fast inference thanks to the transformers and pytorch stack, with weights stored in safetensors format for reduced loading overhead.
• Compatible with Hugging Face text‑classification pipelines, enabling one‑line inference in Python, JavaScript, or Rust.

The architecture mirrors the original BERT‑XL model: 12 transformer layers, 768 hidden dimensions, 12 attention heads, and a classification head (a dense layer with softmax over three classes). The fine‑tuning process retained the pre‑trained language understanding while adapting the final layer to the inclusive‑language task.

Intended use cases revolve around any Italian‑language communication channel that must adhere to inclusive language policies—formal emails, public‑sector documentation, educational material, and corporate communication platforms. The model can be deployed as a backend service, integrated into content‑management systems, or used in real‑time writing assistants to flag or suggest inclusive alternatives.

Benchmark Performance

For a language‑specific classifier, the most relevant benchmarks are accuracy and per‑class F1‑score, which together capture both overall correctness and the balance between precision and recall for each label. The README reports results on a held‑out test set of 1 072 sentences:

The model outperforms both classical baselines (TF‑IDF + MLP/SVM/GB) and a multilingual BERT‑based classifier, achieving the highest overall accuracy (0.89) and a notably strong F1‑score for the “not_inclusive” class (0.92). These metrics demonstrate that the model reliably distinguishes inclusive from non‑inclusive language while correctly ignoring irrelevant sentences—a critical requirement for automated editorial workflows.

Hardware Requirements

Because the model is a full‑size BERT‑XL (≈ 335 M parameters) stored as safetensors, inference memory consumption depends on batch size and sequence length. For a single‑sentence inference (batch = 1, max_length = 128):

• VRAM: ~2 GB of GPU memory is sufficient for the forward pass; a safety margin of 3 GB is recommended to accommodate the tokenizer and PyTorch overhead.
• Recommended GPU: NVIDIA RTX 3060 (12 GB) or any modern GPU with ≥ 4 GB VRAM. For high‑throughput batch processing, a GPU with ≥ 8 GB (e.g., RTX 3070, A100) will keep latency below 30 ms per sentence.
• CPU: A modern 8‑core CPU (e.g., Intel i7‑12700K or AMD Ryzen 7 5800X) can run the model on‑device using the torch‑cpu backend, though inference will be slower (~150 ms per sentence).
• Storage: The model files (weights + tokenizer) occupy roughly 1.2 GB when stored in safetensors format. Adding the README and config files brings the total to ~1.3 GB.
• Performance: On a RTX 3060, batch‑size = 32 yields ~350 tokens / second, which is ample for real‑time applications such as chat‑bots or document‑review pipelines.

Use Cases

The primary intent of the inclusively‑classification model is to support organizations that need to enforce inclusive language standards in Italian. Typical scenarios include:

• Public‑sector communications: Government agencies can automatically scan press releases, policy documents, and citizen‑facing portals for non‑inclusive phrasing.
• Corporate HR and internal communications: Companies can embed the classifier in email clients or intranet editors to flag gender‑biased or exclusionary terminology before it reaches employees.
• Educational material review: Schools and universities can audit textbooks, exam statements, and lecture slides to ensure they meet inclusive language guidelines.
• Content‑management systems: CMS plugins can run the model on article drafts, providing real‑time feedback to authors.
• AI‑assisted writing assistants: Integrated into IDE‑like tools (e.g., Grammarly‑style extensions) to suggest inclusive alternatives on the fly.

Because the model outputs a “not_pertinent” label, it can be safely chained with other classifiers (e.g., topic detection) without generating false positives on unrelated text.

Training Details

The model was fine‑tuned from the Italian BERT‑XL using a supervised classification head. Training data consisted of:

• 8 580 annotated sentences for training.
• 1 073 sentences for validation.
• 1 072 sentences for testing.

All sentences were manually labeled by inclusive‑language experts, ensuring high‑quality ground truth. The fine‑tuning hyper‑parameters were:

• Maximum sequence length: 128 tokens.
• Batch size: 128.
• Learning rate: 5 × 10⁻⁵ (AdamW optimizer).
• Warm‑up steps: 500.
• Number of epochs: 10 (the best checkpoint selected by validation accuracy).

Training was performed on a single NVIDIA A100 GPU (40 GB VRAM) for roughly 3 hours, reflecting the modest dataset size and the efficiency of the AdamW optimizer. The resulting checkpoint was exported in safetensors format for fast loading and reduced memory footprint. p>

The model remains fully fine‑tunable: users can continue training on domain‑specific corpora (e.g., legal or medical Italian) by loading the checkpoint and adjusting the learning rate schedule. Because the base BERT weights are unchanged, further fine‑tuning preserves the rich linguistic knowledge of the original model.

Related Papers

The model’s research foundation is documented in two peer‑reviewed papers:

• Greco et al., “Towards AI‑Assisted Inclusive Language Writing in Italian Formal Communications” (2025) – introduces the inclusive‑language task, dataset, and the BERT‑based approach.
• La Quatra et al., “Inclusively: An AI‑Based Assistant for Inclusive Writing” (2023) – presents a demo system that leverages the classifier in a real‑world writing assistant.

Both papers provide theoretical justification, annotation guidelines, and experimental comparisons that underpin the model’s architecture and training methodology.

Licensing Information

The model is released under the Creative Commons Attribution‑NonCommercial‑ShareAlike 4.0 (CC‑BY‑NC‑SA 4.0) license, as indicated in the README. This license permits:

• Free use, distribution, and modification of the model for non‑commercial purposes.
• Requirement to give appropriate credit to the authors (E‑MIMIC) and the original paper(s).
• Obligation to share any derivative works under the same CC‑BY‑NC‑SA 4.0 license.

Because the license explicitly forbids commercial exploitation, the model < be used in for‑making products, paid SaaS platforms, or any revenue‑generating service without obtaining a separate commercial license from the authors. Academic research, internal tooling, open‑source projects, and educational use are fully allowed, provided attribution is maintained.

If you plan to integrate the model into a commercial workflow, you should contact the authors (E‑MIMIC) to negotiate a commercial agreement or explore alternative models with permissive licenses.