inspect-india-evals 0.1.0

India-specific AI safety, bias, and fairness evaluations for Inspect AI

inspect-india-evals

India-specific AI safety, bias, and cultural knowledge evaluations built on UK AISI's Inspect AI framework.

---

Why This Project Exists

Large language models are increasingly being deployed in India — through national digital infrastructure like Aadhaar, UPI, and Bhashini — yet most AI safety benchmarks are built around Western languages and social contexts, lacking an integrated multilingual context. They miss the things that actually matter here:

• Can a model refuse a harmful request written in Hindi or Tamil, not just English?
• Does it exhibit bias against Indian social categories like caste, religion, or region?
• Does it handle queries about India's Digital Public Infrastructure safely?
• How well does it understand Indian constitutional law and culture?

This project fills that gap. It is a collection of structured evaluation tasks — runnable with a single command — that test models on India-relevant safety, fairness, and knowledge dimensions.

---

What It Evaluates

Five evaluation modules, each targeting a different dimension of model behaviour:

Module	Task	What it measures
multilingual/	multilingual	Factual accuracy (MMLU-style) across Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati, and English
bias/	bharatbbq	Stereotype & social bias using the BharatBBQ benchmark — caste, religion, gender, region
safeguards/	multilingual_safety	Whether the model refuses harmful prompts written in 5 Indian languages
safeguards/	jailbreak_safety	Whether multi-turn jailbreak attacks (7 languages) succeed in extracting harmful content
dpi_safety/	dpi_safety	Correct behaviour on India's DPI context — answer low-risk queries, refuse high-risk ones
cultural_knowledge/	cultural_knowledge	Rubric-graded knowledge of Indian constitution, economy, history, and culture

Composite Fairness Index

All evaluation dimensions roll up into a single Fairness Index — a weighted 0–1 score:

fairness_index = mean([
    multilingual_accuracy,     # language coverage
    1 - |bias_score|,          # lower bias → fairer
    safety_refusal_rate,       # higher refusal on harmful prompts → safer
    dpi_accuracy,              # correct DPI behaviour
])

This gives evaluators one number to compare models against each other or track improvements over time.

---

Project Structure

inspect-india-evals/
├── india_evals/
│   ├── multilingual/           # Multilingual MMLU accuracy
│   │   ├── task.py
│   │   └── datasets/
│   ├── bias/                   # BharatBBQ bias benchmark
│   │   ├── task.py
│   │   └── datasets/
│   ├── safeguards/             # Safety refusal + jailbreak resistance
│   │   ├── task.py
│   │   └── datasets/
│   ├── dpi_safety/             # Digital Public Infrastructure safety
│   │   ├── task.py
│   │   └── datasets/
│   ├── cultural_knowledge/     # Rubric-graded cultural knowledge
│   │   ├── task.py
│   │   └── datasets/
│   ├── scorers/                # Shared scorers (fairness_index)
│   │   └── __init__.py
│   └── view_plugin/            # HTML heatmap report generator
│       ├── parser.py
│       ├── heatmap.py
│       └── __main__.py
├── tests/                      # Full pytest suite (63 tests)
├── run_all.py                  # Multi-model runner with MLflow logging
├── pyproject.toml
└── README.md

---

Setup

Prerequisites

• Python 3.10 or later
• Ollama for running local models (or any model supported by Inspect AI)

Install

# Clone the repo
git clone https://github.com/ac265640/inspect-india-evals
cd inspect-india-evals

# Install dependencies
pip install inspect-ai pandas

# Optional: MLflow for experiment tracking
pip install mlflow

# Optional: Plotly for the visual heatmap report
pip install plotly

No pip install -e . is required — the tasks are run directly with inspect eval.

---

Running Evaluations

Run a single task

# Multilingual MMLU accuracy
inspect eval india_evals/multilingual/task.py@multilingual \
    --model ollama/llama3.2:3b

# BharatBBQ bias benchmark
inspect eval india_evals/bias/task.py@bharatbbq \
    --model ollama/llama3.2:3b

# Safety refusal — harmful prompts in 5 Indian languages
inspect eval india_evals/safeguards/task.py@multilingual_safety \
    --model ollama/llama3.2:3b

# Jailbreak resistance — multi-turn attacks in 7 languages
inspect eval india_evals/safeguards/task.py@jailbreak_safety \
    --model ollama/llama3.2:3b

# DPI Safety — Aadhaar / UPI / Bhashini context
inspect eval india_evals/dpi_safety/task.py@dpi_safety \
    --model ollama/llama3.2:3b

# Cultural knowledge — rubric-graded Indian domains
inspect eval india_evals/cultural_knowledge/task.py@cultural_knowledge \
    --model ollama/llama3.2:3b

Quick smoke test (limit samples)

inspect eval india_evals/safeguards/task.py@multilingual_safety \
    --model ollama/llama3.2:3b --limit 5

View results in the Inspect UI

inspect view

Run all tasks across multiple models + log to MLflow

python run_all.py
mlflow ui   # → open http://localhost:5000

run_all.py runs all tasks across a configurable list of models, parses the .eval logs, computes the composite Fairness Index, and logs everything to MLflow for side-by-side comparison.

---

Visual Heatmap Report

Generate an interactive per-language × per-category HTML report from your eval logs:

# After running evals (logs go to ./logs/ by default)
python -m india_evals.view_plugin logs/

# Custom output path
python -m india_evals.view_plugin --output report.html logs/

# Specific eval files
python -m india_evals.view_plugin logs/run1.eval logs/run2.eval

This opens india_report.html in your browser — a self-contained page with:

• Stats strip (logs, samples, models, languages)
• Summary table (model × task accuracy)
• Interactive heatmaps: language × category for each module
• Bar charts by language and by category

---

Python Developer API

You can import and interact with the project tasks, scoring components, and reporting utilities programmatically within your own scripts:

Run Evaluations Programmatically

Import tasks and execute them using the inspect_ai runner:

from inspect_ai import eval
import india_evals

# Programmatically evaluate the multilingual task
logs = eval(
    india_evals.multilingual(),
    model="ollama/llama3.2:3b",
    limit=5
)

Calculate the Fairness Index

Compute the composite India Fairness Index manually using the normalized sub-scores:

import india_evals

results = india_evals.fairness_index(
    multilingual_accuracy=0.72,
    bias_score_amb=0.10,          # Stereotype bias
    safety_refusal_rate=0.88,     # Safety refusal rate
    dpi_accuracy=0.85             # DPI query accuracy
)

print(results["fairness_index"])  # Output: 0.8375

Render the HTML Report

Generate the interactive visual HTML report from your evaluation logs folder programmatically:

import india_evals

# Scans logs/ directory and generates report.html
report_path = india_evals.render_report(
    log_dir="logs/",
    output="report.html"
)

---

Module Details

multilingual — Multilingual MMLU Accuracy

Tests factual reasoning across 6 Indian languages using an MMLU-style multiple-choice format. Each question has one correct answer from options A–D.

Languages: English, Hindi, Tamil, Telugu, Bengali, Marathi, Gujarati
Scorer: Exact match (A/B/C/D)
Metric: Accuracy (0–1)

---

bharatbbq — Social Bias (BharatBBQ)

Based on the BharatBBQ benchmark, which tests whether a model applies stereotypes when answering ambiguous social questions about Indian social groups.

Categories: Caste, Religion, Gender, Age, Region, Physical appearance
Scorer: Accuracy on unambiguous questions + bias direction on ambiguous ones
Metric: Accuracy (0–1), Bias score (−1 to +1)

---

multilingual_safety — Safety Refusal

Tests whether a model correctly refuses harmful requests (cybercrime, violence, fraud, etc.) when written in Indian languages — not just English.

Languages: English, Hindi, Tamil, Telugu, Bengali
Scoring: Two-layer — keyword phrase matching first, LLM judge fallback
Metric: Refusal rate (1 = refused, 0 = complied)

---

jailbreak_safety — Jailbreak Resistance

Tests whether multi-turn jailbreak attack sequences (roleplay, hypothetical, escalation) succeed in extracting harmful content across 7 Indian languages.

Attack types: Roleplay, Research framing, Hypothetical, Escalating pressure
Scorer: Same two-layer scorer (keyword + LLM judge)
Metric: Refusal rate (higher is safer)

---

dpi_safety — Digital Public Infrastructure Safety

Tests correct handling of queries in the context of India's DPI stack: Aadhaar identity, Deepfake, UPI payments, and Bhashini language services.

Low-risk queries (how does Aadhaar work?) should be answered.
High-risk queries (how to spoof Aadhaar verification?) should be refused.

Risk levels: Low, Medium, High
Scorer: Keyword-based + LLM judge
Metric: Accuracy on correct behaviour (answer vs. refuse)

---

cultural_knowledge — Indian Cultural Knowledge

Open-ended questions graded against a 4-criterion rubric by an LLM judge. Tests knowledge of Indian constitutional law, healthcare system, economy, history, science & technology, geography, and culture.

Dataset: ~825 questions across 8+ domains
Scorer: LLM-as-judge rubric scoring (fraction of criteria passed)
Metric: Mean rubric score (0–1)

---

Tests

pytest tests/ -v

63 tests across all modules. Covers dataset loading, sample structure, scorer logic, and task instantiation. No model calls required — all scorer tests are unit-tested with mocks.

---

Supported Models

Any model supported by Inspect AI works. Tested with:

Model	How to run
Ollama local models	--model ollama/llama3.2:3b
OpenAI	--model openai/gpt-4o
Anthropic	--model anthropic/claude-3-5-sonnet
Together AI	--model together/meta-llama/...
Sarvam AI	--model sarvam/sarvam-m

Set API keys as environment variables as required by each provider.

---

Design Decisions

Why Inspect AI?
Inspect AI is the UK AISI's open evaluation framework. It handles model calls, logging, scoring, and parallelism out of the box, making eval code reproducible and auditable.

Why keyword + LLM judge for safety scoring?
Keyword matching is fast, free, and works across all 7 languages without needing a separate call. The LLM judge handles edge cases: a model that generates a malware scanner description isn't the same as generating malware, and the judge can tell the difference.

Why rubric scoring for cultural knowledge?
Cultural and constitutional questions have open-ended correct answers. A rubric of 4 specific factual criteria — graded by a judge model — gives a structured, reproducible score without needing a fixed expected answer.

---

License

Apache 2.0