vikhram-labs/NariRaksha-1K

NariRaksha-100K

A structured dataset of Indian women's safety scenarios paired with risk classification, step-by-step reasoning, recommended actions, and applicable legal context (BNS / IT Act / PWDVA), built to support fine-tuning of safety-assessment language models such as NariRaksha-3B.

Dataset Summary

Each row describes a safety scenario and a corresponding structured assessment: risk category, severity, reasoning, a recommended action for the victim, and the relevant Indian legal provisions. The dataset is intended as training/fine-tuning data for models that triage or explain safety situations — not as a verified legal, medical, or emergency-response reference (see Limitations below).

Dataset Structure

Data Fields

Field	Type	Description
scenario	string	Narrative description of the safety scenario
language	string	Language of the scenario text
risk_type	string	Risk category (e.g. coercive control, cyberstalking)
severity	string	low / medium / high / critical
reasoning	string	Step-by-step rationale for the classification
recommended_action	string	Suggested action for the victim, often including a helpline
legal_context	string	Applicable BNS / IT Act / PWDVA sections
confidence	float64	Model/annotator-assigned confidence score (0–1)

Data Splits

A single split of approximately 100,000 rows. No predefined train/validation/test split is provided — downstream users should partition according to their own evaluation protocol (e.g. stratified by risk_type and severity to preserve class balance).

Example Row

{
  "scenario": "A 25-year-old woman from Hyderabad reported an incident of coercive control. The perpetrator had been targeting her for several weeks, causing high-level distress. She sought help after the situation escalated and began affecting her professional life.",
  "language": "English",
  "risk_type": "coercive control",
  "severity": "high",
  "reasoning": "This case constitutes coercive control based on the sustained, unwanted conduct directed at the victim. The high severity is consistent with the documented impact on her daily functioning and emotional well-being. The pattern of behaviour meets the legal threshold under Indian law.",
  "recommended_action": "Contact Shakti Shalini (011-24373736) or the National Commission for Women (7217735372). Ensure personal safety and avoid engaging with the perpetrator.",
  "legal_context": "BNS Section 85 (cruelty by husband/relatives), PWDVA Section 3",
  "confidence": 0.88
}

Dataset Creation

Source Data

The dataset is a mix of synthetic, LLM-generated scenarios and real-derived templates — some scenarios are modeled on patterns from anonymized, publicly reported case types and then rewritten/generalized; others are fully synthetic. No raw real-world case records, personal data, or identifying information are included.

Generation Process

reasoning and recommended_action fields were generated using a template-conditioned approach keyed largely on risk_type and severity combinations, rather than being independently authored per scenario. As a result, a substantial share of rows share near-identical boilerplate text in these fields (see Dataset Composition Notes below) — this is a property of the generation pipeline, not a data quality defect, but it should inform how you sample, deduplicate, or weight the data during training.

Dataset Composition Notes

Value distributions are imbalanced across categorical fields. For example:

• risk_type: no single category dominates, but coverage across categories is uneven (e.g. coercive control ≈ 7.9% of rows).
• severity: high ≈ 24.5% of rows, with the remaining mass spread across low / medium / critical.
• reasoning: the single most common value covers ≈ 31% of all rows.
• recommended_action: the single most common value covers ≈ 33.9% of all rows.

This means reasoning and recommended_action are far less diverse than scenario text. Users training generative models on this data should be aware that naive fine-tuning may cause a model to overfit to a small set of stock phrases for these fields.

Limitations and Known Issues

• Legal and helpline content is only partially verified. Citations to BNS sections, IT Act provisions, PWDVA sections, and helpline numbers are a mix of manually verified and LLM-generated, unverified entries. Some legal references or contact numbers in this dataset may be inaccurate, outdated, or incorrectly matched to the scenario. Do not treat any single row as a verified legal or emergency-response source.
• Synthetic content. A portion of scenarios are fully synthetic and do not describe real events. They are illustrative, not statistically representative of actual incident rates or demographics.
• Templated reasoning/actions. As noted above, reasoning and recommended_action exhibit low lexical diversity relative to scenario.
• Not for deployment as-is. Models trained on this data for real-world safety triage, legal advice, or emergency response should undergo independent legal review and human-in-the-loop validation before any operational use. This dataset is a research/fine-tuning resource, not a certified safety or legal database.

Considerations for Using the Data

This dataset deals with sensitive subject matter (gender-based violence, stalking, coercive control, and related harms). Users should:

• Avoid presenting model outputs trained on this data as authoritative legal or medical advice.
• Independently verify any helpline number or legal citation before surfacing it to an end user.
• Be mindful of language and cultural context when adapting scenarios beyond the Indian legal system (BNS / IT Act / PWDVA are India-specific).

Citation

If you use this dataset, please cite the associated casebook chapter:

@incollection{nariraksha2026,
  title     = {NariRaksha: Gender-Responsive AI for Women's Safety},
  author    = {Vikhram S and Jeffin Gracewell J},
  booktitle = {India AI Impact Summit 2026 Casebook on AI and Gender Empowerment},
  year      = {2026},
  publisher = {Ministry of Electronics and Information Technology (MeitY), Government of India and UN Women}
}

License

Apache 2.0. See the LICENSE file for details.