---

language:language:

• en- en

library_name: transformerslibrary_name: transformers

pipeline_tag: token-classificationpipeline_tag: token-classification

language: - en library_name: transformers pipeline_tag: token-classification tags: - ner - legal - indian-law - bert license: cc-by-nc-4.0 model-index: - name: IN_Lexi_X_BERT results: - task: name: Named Entity Recognition type: token-classification dataset: name: Indian Legal NER (enriched) type: custom split: test metrics: - name: Micro F1 type: f1 value: 0.7773406766325727 - name: Macro F1 type: f1 value: 0.780584439098286

IN_Lexi_X_BERT — Indian Legal NER (Improved)

This repository contains a fine-tuned BERT model (InLegalBERT-style) for Named Entity Recognition on Indian legal texts.

• Task: Token classification (NER)
• Domain: Indian legal documents
• Max sequence length: 512
• Framework: Hugging Face Transformers

What’s new

This version improves NER performance on the test split compared to the previous push.

• Micro F1: 0.77734
• Macro F1: 0.78058

Per-class highlights (F1): COURT ≈ 0.894, PROVISION ≈ 0.939, STATUTE ≈ 0.925. See compare_checkpoints_report.json for full details.

Labels

The model predicts BIO-formatted legal entities:

• B-CASE_NUMBER, I-CASE_NUMBER
• B-COURT, I-COURT
• B-DATE, I-DATE
• B-GPE, I-GPE
• B-JUDGE, I-JUDGE
• B-LAWYER, I-LAWYER
• B-ORG, I-ORG
• B-OTHER_PERSON, I-OTHER_PERSON
• B-PETITIONER, I-PETITIONER
• B-PRECEDENT, I-PRECEDENT
• B-PROVISION, I-PROVISION
• B-RESPONDENT, I-RESPONDENT
• B-STATUTE, I-STATUTE
• B-WITNESS, I-WITNESS
• O

These correspond to the config.json id2label mapping packaged with the checkpoint.

Quick start

from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline

repo_id = "shreyas2809/IN_Lexi_X_BERT"

tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForTokenClassification.from_pretrained(repo_id)

ner = pipeline(
    "token-classification",
    model=model,
    tokenizer=tokenizer,
    aggregation_strategy="simple",  # aggregates B/I into spans
)

text = "In The High Court Of Kerala At Ernakulam ..."
print(ner(text))

Intended use

• Information extraction from Indian legal case texts
• Downstream legal analytics (parties, courts, statutes, provisions, etc.)

Evaluation

The provided metrics are computed on a held-out test split of the Indian Legal NER dataset (enriched). For a more detailed analysis, generate a confusion matrix and per-class metrics locally.

Limitations

• Domain-specific: optimized for Indian legal language patterns
• Long documents: sequences >512 tokens are truncated
• Class imbalance: some labels are less frequent relative to 'O'

License

CC-BY-NC 4.0

• B-ORG, I-ORG model=model,

• B-OTHER_PERSON, I-OTHER_PERSON tokenizer=tokenizer,

• B-PETITIONER, I-PETITIONER aggregation_strategy="simple", # aggregates B/I into single spans

• B-PRECEDENT, I-PRECEDENT)

• B-PROVISION, I-PROVISION

• B-RESPONDENT, I-RESPONDENTtext = "On query by the Bench about an entry of Rs. 1,31,37,500 on deposit side of Hongkong Bank account of ..."

• B-STATUTE, I-STATUTEprint(ner(text))

• B-WITNESS, I-WITNESS```

• O

Example output (structure):

These correspond to the config.json id2label mapping packaged with the checkpoint.

## Quick start[

```python  {

from transformers import AutoTokenizer, AutoModelForTokenClassification, pipeline    'entity_group': 'ORG',

    'score': 0.98,

repo_id = "shreyas2809/IN_Lexi_X_BERT"    'word': 'Hongkong Bank',

    'start': 86,

tokenizer = AutoTokenizer.from_pretrained(repo_id)    'end': 99,

model = AutoModelForTokenClassification.from_pretrained(repo_id)  },

  # ...

ner = pipeline(]

    "token-classification",```

    model=model,

    tokenizer=tokenizer,## Intended use

    aggregation_strategy="simple",  # aggregates B/I into spans- Information extraction from Indian legal case texts

)- Downstream legal analytics (parties, courts, statutes, provisions, etc.)



text = "In The High Court Of Kerala At Ernakulam ..."## Training and data

print(ner(text))The model was fine-tuned on Indian legal NER data prepared by the author. BIO-formatted splits are not included here; labels are baked into the checkpoint. If you want to reproduce evaluation locally, run a confusion-matrix evaluation on your BIO files using a simple script like:

## Intended use# local usage example (not part of the Hub repo)

- Information extraction from Indian legal case textspython evaluate_confusion_matrix.py \

- Downstream legal analytics (parties, courts, statutes, provisions, etc.)  --model-dir best \

  --data-path enriched_data/test.bio \

## Evaluation  --output-image reports/test_confusion.png

The provided metrics are computed on a held-out test split of the Indian Legal NER dataset (enriched). For a more detailed analysis, generate a confusion matrix and per-class metrics locally.```



## LimitationsThis script aligns word-piece tokens to word labels, computes a confusion matrix (optionally excluding the dominant 'O' class), and can export a heatmap.

- Domain-specific: optimized for Indian legal language patterns

- Long documents: sequences >512 tokens are truncated## Limitations

- Class imbalance: some labels are less frequent relative to 'O'- Domain-specific: optimized for Indian legal language and may generalize poorly to other domains or jurisdictions.

- Long documents: sequences longer than 512 tokens are truncated.

## License- Class imbalance: some entities may be under-represented relative to 'O'.

Not specified.

## Citation
If you use this model in academic or industry work, please cite this repository and the underlying InLegalBERT base.

## License
CC-BY-NC 4.0