README.md

metadata

license: apache-2.0
datasets:
  - ArchitRastogi/Italian-BERT-FineTuning-Embeddings
language:
  - it
metrics:
  - Recall@1
  - Recall@100
  - Recall@1000
  - Average Precision
  - NDCG@10
  - NDCG@100
  - NDCG@1000
  - MRR@10
  - MRR@100
  - MRR@1000
base_model:
  - dbmdz/bert-base-italian-xxl-uncased
new_version: 'true'
pipeline_tag: feature-extraction
library_name: transformers
tags:
  - information-retrieval
  - contrastive-learning
  - embeddings
  - italian
  - fine-tuned
  - bert
  - retrieval-augmented-generation
model-index:
  - name: bert-base-italian-embeddings
    results:
      - task:
          type: information-retrieval
        dataset:
          name: mMARCO
          type: mMARCO
        metrics:
          - name: Recall@1000
            type: Recall
            value: 0.9719
          - name: NDCG@1000
            type: Normalized Discounted Cumulative Gain
            value: 0.4391
          - Average Precision: AP
            type: Precision
            value: 0.3173
        source:
          name: Fine-tuned Italian BERT Model Evaluation
          url: https://github.com/unicamp-dl/mMARCO

bert-base-italian-embeddings: A Fine-Tuned Italian BERT Model for IR and RAG Applications

Model Overview

This model is a fine-tuned version of dbmdz/bert-base-italian-xxl-uncased tailored for Italian language Information Retrieval (IR) and Retrieval-Augmented Generation (RAG) tasks. It leverages contrastive learning to generate high-quality embeddings suitable for both industry and academic applications.

Model Size

• Size: Approximately 450 MB

Training Details

• Base Model: dbmdz/bert-base-italian-xxl-uncased
• Dataset: Italian-BERT-FineTuning-Embeddings
  • Derived from the C4 dataset using sliding window segmentation and in-document sampling.
  • Size: ~5GB (4.5GB train, 0.5GB test)
• Training Configuration:
  • Hardware: NVIDIA A40 GPU
  • Epochs: 3
  • Total Steps: 922,958
  • Training Time: Approximately 5 days, 2 hours, and 23 minutes
• Training Objective: Contrastive Learning

Evaluation Metrics

Evaluations were performed using the mMARCO dataset, a multilingual version of MS MARCO. The model was assessed on 6,980 queries.

Results Comparison

Metric	Base Model (dbmdz/bert-base-italian-xxl-uncased)	facebook/mcontriever-msmarco	Fine-Tuned Model
Recall@1	0.0026	0.0828	0.2106
Recall@100	0.0417	0.5028	0.8356
Recall@1000	0.2061	0.8049	0.9719
Average Precision	0.0050	0.1397	0.3173
NDCG@10	0.0043	0.1591	0.3601
NDCG@100	0.0108	0.2086	0.4218
NDCG@1000	0.0299	0.2454	0.4391
MRR@10	0.0036	0.1299	0.3047
MRR@100	0.0045	0.1385	0.3167
MRR@1000	0.0050	0.1397	0.3173

Note: The fine-tuned model significantly outperforms both the base model and facebook/mcontriever-msmarco across all metrics.

Usage

You can load and use the model directly with the Hugging Face Transformers library:

# Load model directly
from transformers import AutoTokenizer, AutoModelForMaskedLM

tokenizer = AutoTokenizer.from_pretrained("ArchitRastogi/bert-base-italian-embeddings")
model = AutoModelForMaskedLM.from_pretrained("ArchitRastogi/bert-base-italian-embeddings")

# Example usage
text = "Stanchi di non riuscire a trovare il partner perfetto?"
inputs = tokenizer(text, return_tensors="pt")
outputs = model(**inputs)

Intended Use

This model is intended for:

• Information Retrieval (IR): Enhancing search engines and retrieval systems in the Italian language.
• Retrieval-Augmented Generation (RAG): Improving the quality of generated content by providing relevant context.

Suitable for both industry applications and academic research.

Limitations

• The model may inherit biases present in the C4 dataset.
• Performance is primarily evaluated on mMARCO; results may vary with other datasets.

---

Contact

Archit Rastogi
📧 architrastogi20@gmail.com