kabir4756/manas-dataset

Manas AI Training Dataset

This repository contains the official high-quality fine-tuning training dataset for Manas — a warm, culturally-grounded mental health companion AI designed for Indian cultural contexts.

Dataset Overview

The dataset is organized into 11 distinct categories representing psychological pressures and emotional challenges common in urban and semi-urban Indian demographics. Each dialogue is generated using a distinct thematic persona archetype, ensuring representation across age groups, genders, professions, regions, and languages (English, Hindi, and Hinglish).

All conversations follow a strict 5-stage therapeutic arc (Silent Assessment -> Exploration -> Surfacing -> Actionable Culturally Grounded Suggestion -> Space for Closing) and strictly enforce clinical safety and ethical boundaries.

Dataset Splits and Size

• Total Validated Conversations: 137
• Train Split (90%): 123 (ChatML Format)
• Eval Split (10%): 14 (ChatML Format)

Category Breakdown

Category	Description	Total Conversations
career_anxiety	Job loss, corporate pressure, stagnation, ageism	12
parental_expectations	Family business pressure, unwanted marriages, academic comparison	12
relationship_stress	Modern dating, attachment issues, divorce/separation stigma	12
relocation_loneliness	Small town to metro migration, isolation in tech hubs	12
financial_stress	Debt cycles, sandwich generation pressures	10
academic_pressure	Competitive exams, coaching centers, fear of failure	12
identity_confusion	Coming out, gender non-conformance, double life	13
burnout_motivation	Overachiever collapse, loss of meaning/joy	12
grief_loss	Bereavement, relationship ending, loss of physical health	15
coping_mechanisms	Self-harm avoidance, substance reliance, workaholism	13
trauma_patterns	Intergenerational trauma, childhood abuse scars, relationship abuse	14

Format Structure

Conversations are formatted in standard ChatML to align with modern fine-tuning libraries:

{
  "messages": [
    {
      "role": "system",
      "content": "..."
    },
    {
      "role": "user",
      "content": "..."
    },
    {
      "role": "assistant",
      "content": "..."
    }
  ]
}

Validation & Safety Guarantees

All records successfully passed a rigorous automated validation suite ensuring:

• No Diagnostics: Avoids clinical terms (e.g. diagnosing with clinical depression, PTSD).
• No Hollow Reassurance: Free from toxic positivity ("stay positive", "everything happens for a reason").
• Safety Protocol: Multi-turn safety and immediate referral trigger system on acute self-harm signs.
• Single Question Principle: Assistant asks no more than 1 question per turn to prevent overwhelm.
• Role Alternation: Enforces strict role alternation patterns.