You need to agree to share your contact information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

Legal Language Modelling Challenge — CS-781 Capstone

This repository contains the datasets for all seven phases of the Legal Language Modelling Challenge (CS-781 Capstone).

The challenge covers a range of legal NLP tasks, progressing from legal language modelling and information extraction to rhetorical role prediction, case judgement prediction, legal statute identification, legal summarization, and prior case retrieval.

Dataset Overview

Phase Task Train Dev Test
Phase 1 Legal Language Model Pre-training ✓ — ✓
Phase 2 Legal Named Entity Recognition (L-NER) ✓ — ✓
Phase 3 Rhetorical Role Prediction (RR) ✓ ✓ ✓
Phase 4 Case Judgement Prediction (CJPE) ✓ ✓ ✓
Phase 5 Legal Statute Identification (LSI) ✓ ✓ ✓
Phase 6 Legal Summarization ✓ — ✓
Phase 7 Prior Case Retrieval (PCR) ✓ ✓ ✓

Phase 1 — Legal Language Model Pre-training

Task

The objective of this phase is to train a compact legal language model from scratch and generate a continuation for a given legal-text prompt.

Dataset Structure

Phase_1_Pretraining/
├── Train/
│   ├── train_acts.json
│   ├── train_high_courts.json
│   ├── train_supreme_court.json
│   └── train_tribunals.json
│
└── Test/
    └── test.json

Training Data

The training corpus contains more than 4.5 billion tokens of legal text, provided in JSON format.

The training files contain legal text records.

Test Data

Each test example consists of a 512-word legal-text prompt:

{
  "1": {
    "text": "The issue involved in these petitions is as to whether ..."
  },
  "2": {
    "text": "In ..."
  }
}

The model must generate the next 128 words for every test prompt.

Expected Submission Format

The prediction file must be named prediction.json and contain one prediction for every test ID:

{
  "1": {
    "text": "Generated 128-word continuation ..."
  },
  "2": {
    "text": "Generated 128-word continuation ..."
  }
}

The text field must contain only the generated continuation and must not include the original 512-word prompt.

Evaluation

  • Primary: BERTScore F1
  • Secondary: ROUGE-L F1, BLEU-4

Phase 2 — Legal Named Entity Recognition (L-NER)

Task

The objective of this phase is to perform named entity recognition over Indian legal documents using character-level entity spans.

Dataset Structure

Phase_2_LNER/
├── Train/
│   ├── fold1-text.json
│   ├── fold1-gold.json
│   ├── fold2-text.json
│   ├── fold2-gold.json
│   └── labels.txt
│
└── Test/
    └── fold3-text.json

Training Data

The training text files contain document IDs mapped to their complete document text:

{
  "115651329": "REPORTABLE IN THE SUPREME COURT OF INDIA ..."
}

The training gold files contain the corresponding entity annotations:

{
  "115651329": [
    [137, 153, "RESP", "STATE OF HARYANA"],
    [252, 261, "DATE", "19.8.2011"],
    [276, 308, "COURT", "High Court of Punjab and Haryana"]
  ]
}

Each annotation follows:

[start_character, end_character, label, entity_text]

The labels.txt file contains the permitted entity labels.

Test Data

The test data contains only the documents:

{
  "1128835": "REPORTABLE IN THE SUPREME COURT OF INDIA ..."
}

The corresponding entity annotations are private.

Expected Submission Format

Predictions are provided as a JSON object keyed by document ID:

{
  "1128835": [
    [18, 40, "COURT"],
    [114, 162, "CASENO"],
    [213, 229, "RESP"]
  ]
}

Each prediction must contain:

[start_character, end_character, label]

An optional fourth field containing the entity text is also permitted:

[start_character, end_character, label, entity_text]

The entity text is not used for scoring. The character span and entity label must exactly match the private annotation.

Evaluation

  • Primary: Strict Macro-F1 across the 12 entity labels
  • Secondary: Strict Macro-Precision, Strict Macro-Recall

Phase 3 — Rhetorical Role Prediction (RR)

Task

The objective of this phase is to predict the rhetorical role of each sentence in a legal document.

The dataset covers two legal domains:

  • CL — Corporate Law
  • IT — Information Technology

Dataset Structure

Phase_3_RR/
├── Train/
│   ├── CL_train.json
│   ├── CL_dev.json
│   ├── IT_train.json
│   ├── IT_dev.json
│   └── label_vocab.json
│
└── Test/
    ├── CL_test.json
    └── IT_test.json

Training Data

Each document contains a list of sentences and a corresponding list of rhetorical-role labels:

{
  "id": "CCI_Noida_Software_Technology_Park_Ltd_vs_Star_India...",
  "text": [
    "Noida Software Technology Park Limited ...",
    "The Commission observed that ..."
  ],
  "labels": [
    "Fact",
    "RulingByPresentCourt"
  ]
}

There is a one-to-one correspondence between text and labels. That is:

len(text) == len(labels)

and the label at position i corresponds to the sentence at position i.

label_vocab.json contains the permitted rhetorical-role labels.

Test Data

The test data contains only the sentences:

{
  "HC_Raghupati_Singhania__vs__Competition_Commission...": [
    "Manmohan, J. 1 .",
    "Present writ petition has been filed, challenging the notice dated ...",
    "...",
    "In the event that such an application is filed, it would be decided ..."
  ]
}

The corresponding labels are private.

Expected Submission Format

Predictions must contain one label for every test sentence, in the same order as the input sentences:

{
  "HC_Raghupati_Singhania__vs__Competition_Commission...": [
    "Issue",
    "Fact",
    "Fact",
    "RulingByPresentCourt"
  ]
}

The number of predictions must exactly match the number of sentences in the corresponding test document. Every prediction must be one of the permitted labels in label_vocab.json.

Evaluation

  • Primary: Macro-F1 across the 13 rhetorical-role labels
  • Secondary: Macro-Precision, Macro-Recall

Phase 4 — Case Judgement Prediction (CJPE)

Task

The objective of this phase is to predict the binary judgement outcome for each test case.

Dataset Structure

The source data contains separate single-instance and multi-instance training/development splits. These were combined within each corresponding split during preprocessing.

Phase_4_CJPE/
├── Train/
│   ├── train.json
│   └── dev.json
│
└── Test/
    └── test.json

Training Data

The training data in train.json contains records with:

{
  "id": "...",
  "text": "...",
  "label": 1
}

The training set combines the original single-instance and multi-instance training data.

Development Data

The development data in dev.json contains records with:

{
  "id": "...",
  "text": "...",
  "label": 0
}

Test Data

The test data in test.json contains:

{
  "id": "...",
  "text": "..."
}

The output/label is intentionally removed from the test data.

Expected Submission Format

Predictions are provided in the following form:

[
  {"id": "...", "pred": 0},
  {"id": "...", "pred": 1}
]

where pred is the binary judgement prediction.

Evaluation

  • Primary: Macro-F1
  • Secondary: Precision, Recall

Phase 5 — Legal Statute Identification (LSI)

Task

The objective of this phase is to identify the applicable legal statute(s) for each document.

Dataset Structure

Phase_5_LSI/
├── Train/
│   ├── train.json
│   └── dev.json
│
└── Test/
    └── test.json

Training Data

The training and development data in train.json and dev.json contain records with:

{
  "id": "...",
  "text": "...",
  "labels": [1, 7, 23]
}

labels contains the applicable statute IDs for the example.

Test Data

The test data in test.json contains:

{
  "id": "...",
  "text": "..."
}

The statute labels are removed from the test set.

Expected Submission Format

Each prediction is provided in the following form:

[
  {
    "id": "...",
    "pred": [1, 7, 23]
  }
]

pred contains the applicable statute IDs. The statute IDs must be valid and unique for a given example.

Evaluation

  • Primary: Macro-F1
  • Secondary: Micro-F1, Subset Accuracy

Phase 6 — Legal Summarization

Task

The objective of this phase is to generate a legal summary from the corresponding legal document.

Dataset Structure

Phase_6_Summ/
├── Train/
│   └── train.json
│
└── Test/
    └── test.json

Training Data

The training data in train.json contains records with:

{
  "id": "...",
  "document": "...",
  "summary": "..."
}

Test Data

The test data in test.json contains:

{
  "id": "...",
  "document": "..."
}

Expected Submission Format

Predictions are provided in the following form:

[
  {
    "id": "...",
    "summary": "Generated legal summary ..."
  }
]

Evaluation

  • Primary: BERTScore F1
  • Secondary: ROUGE-1 F1, ROUGE-L F1

Phase 7 — Prior Case Retrieval (PCR)

Task

The objective of this phase is to retrieve relevant prior legal cases from a candidate pool for each query case.

This phase is a retrieval task. For each query case, the model must rank candidate case documents by relevance and return the most relevant candidate case IDs.

Dataset Structure

Each split contains a query file and a corresponding candidate-pool file. The candidate pools are split-specific and should be used with the corresponding query split.

Phase_7_PCR/
├── Train/
│   ├── train.json
│   └── train_candidates.json
│
├── Dev/
│   ├── dev.json
│   └── dev_candidates.json
│
└── Test/
    ├── test.json
    └── test_candidates.json

The dataset contains:

Split Queries Candidates
Train 827 4,320
Dev 118 1,023
Test 237 1,727

Training Data

The training query file contains query IDs mapped to the query text and its relevant candidate case IDs:

{
  "100120460": {
    "text": [
      "CASE NO.: Appeal (civil) 2387 of 2001 ...",
      "THOMAS, J.",
      "..."
    ],
    "relevant_candidates": [
      "0000586994",
      "0000772259",
      "0001182839"
    ]
  }
}

The corresponding train_candidates.json contains the candidate case documents:

{
  "0000586994": {
    "text": [
      "JUDGMENT ...",
      "..."
    ]
  }
}

The query and candidate IDs are IndianKanoon case IDs. The text field contains the case document divided into sentences.

Development Data

The development data follows the same structure as the training data:

{
  "100120460": {
    "text": [
      "Query case text ..."
    ],
    "relevant_candidates": [
      "0000586994",
      "0000772259"
    ]
  }
}

dev_candidates.json contains the candidate case documents available for the development queries.

The development relevance labels can be used for local evaluation.

Test Data

The test query file contains the query case documents without the relevance labels:

{
  "0000021652": {
    "text": [
      "PETITIONER: <NAME> Vs. RESPONDENT: <NAME> & ORS. ...",
      "..."
    ]
  }
}

The corresponding candidate documents are provided in test_candidates.json:

{
  "0000000987": {
    "text": [
      "JUDGMENT <NAME>, J. ...",
      "..."
    ]
  }
}

The relevant_candidates field is intentionally removed from the public test queries. The corresponding relevance labels are private.

Only candidates from test_candidates.json should be retrieved for the test queries.

Expected Submission Format

The prediction file must be named prediction.json and contain one ranked list of candidate IDs for every test query:

{
  "0000021652": [
    "0000590455",
    "0000916840",
    "0001481813",
    "0001907199",
    "0000000987",
    "0000001566",
    "0000002886"
  ],
  "0000030721": [
    "0000004354",
    "0000320809",
    "0000516439",
    "0000611866",
    "0000685234",
    "0000785119",
    "0000849101"
  ]
}

Each query must contain exactly 7 unique candidate IDs, ranked from highest to lowest relevance. All submitted candidate IDs must belong to the corresponding test candidate pool.

Only the top 7 retrieved candidates are evaluated.

Evaluation

  • Primary: Micro-F1@7
  • Secondary: Micro-Precision@7, Micro-Recall@7

The gold relevance sets may contain fewer or more than 7 relevant candidates. They are not truncated or padded. The @7 metric evaluates only the first 7 retrieved candidates against the complete set of relevant candidates for each query.


Submission Structure

For the competition, submissions are packaged as:

submission.zip
└── prediction.json

The prediction.json structure differs by phase, as described above. The prediction file must correspond exactly to the IDs and structural requirements of the relevant test set.

Phase Summary

Phase Task Input Output
1 Legal Language Model 512-word legal prompt 128-word continuation
2 L-NER Legal document Character spans + entity labels
3 Rhetorical Role Legal sentences One rhetorical-role label per sentence
4 CJPE Legal case text Binary judgement prediction
5 LSI Legal document Applicable statute ID(s)
6 Summarization Legal document Legal summary
7 Prior Case Retrieval Query case + candidate case pool Ranked list of 7 candidate case IDs

Citation / Source

The datasets are used as part of the Legal Language Modelling Challenge — CS-781 Capstone.

Downloads last month
73