KiranaChain is a large-scale benchmark dataset designed to study decentralized retail negotiation dynamics, informal credit allocation mechanisms, supply-chain disruptions, and multilingual bargaining behavior within Indian micro-retail ecosystems. The dataset captures structured negotiation trajectories between wholesale distributors and independent Kirana retailers, providing a unified representation of financial decision-making, logistics constraints, environmental volatility, and conversational interaction.

The benchmark contains 1,000,000 negotiation-turn observations organized into 125,000 complete negotiation trajectories, each consisting of eight sequential bargaining stages. Every observation is represented through a 40-dimensional structured state space integrating agent-level attributes, credit risk indicators, market signals, supply-chain conditions, price evolution dynamics, and multilingual dialogue telemetry.

KiranaChain is intended to support research across reinforcement learning, multi-agent systems, graph intelligence, behavioral economics, supply-chain optimization, conversational AI, credit risk modeling, and decision-making under uncertainty.

ARCHITECTURAL OVERVIEW & TAXONOMY

Decentralized Retail Networks as Dynamic Economic Graphs

The Indian Kirana ecosystem represents one of the world's largest decentralized retail infrastructures. Unlike vertically integrated retail chains, Kirana networks operate through highly localized bilateral relationships between independent retailers and distributors. Transactional decisions emerge from repeated interactions shaped by trust, liquidity availability, repayment history, inventory pressure, regional demand fluctuations, and external supply disruptions.

The resulting economic structure can be represented as a dynamic weighted graph:

$G_{t} = (V, E_{t}, W_{t})$

where:

$$(V)$$ denotes retailers and distributors.
$$(E_t)$$ denotes active trading relationships at time (t).
$$(W_t)$$ encodes evolving trust, credit exposure, transaction volume, and repayment behavior.

Unlike centralized commerce systems, this graph continuously evolves through local negotiations, producing highly non-linear behavioral patterns and heterogeneous decision trajectories.

Informal Credit Systems (Udhar)

A defining characteristic of decentralized retail commerce is the prevalence of relationship-based credit arrangements.

Credit allocation depends upon:

Historical repayment consistency.
Transaction frequency.
Outstanding liabilities.
Liquidity stress conditions.
Distributor-specific risk tolerance.

The interaction between trust and credit allocation forms a feedback system:

$L_{t} = f (T_{t}, R_{t}, H_{t})$

where:

$$(L_t)$$ = allocated credit limit
$$(T_t)$$ = trust score
$$(R_t)$$ = repayment velocity
$$(H_t)$$ = historical transaction profile

This mechanism directly influences negotiation outcomes, settlement probability, and long-term network stability.

Supply-Chain Dynamics and Perishable Logistics

Retail negotiations occur within operational environments affected by:

Fuel price fluctuations
Transportation bottlenecks
Inventory scarcity
Seasonal demand shifts
Weather-related disruptions
Commodity perishability

Perishable commodities introduce time-dependent value degradation:

where:

$$(V_0)$$ is initial commodity value.
$$(\delta)$$ is decay rate.
$$(t)$$ denotes storage or transit duration.

Such dynamics substantially influence bargaining strategies and settlement behavior.

Multilingual Negotiation Intelligence

Negotiation trajectories include multilingual conversational interactions reflecting realistic commercial communication patterns.

Dialogue streams encode:

Price concession requests
Inventory concerns
Credit extensions
Relationship maintenance
Settlement commitments
Demand forecasts

The conversational component provides opportunities for:

Dialogue state tracking
Strategic language modeling
Negotiation policy learning
Multilingual conversational intelligence research

Dataset Taxonomy

The 40-feature architecture is organized into four major subsystems:

Subsystem	Columns	Purpose
Agent & Geographic Intelligence	1–8	Identity, location, commodity context
Financial & Credit Intelligence	9–18	Trust, credit exposure, repayment dynamics
Supply-Chain Intelligence	19–28	Environmental and logistics conditions
Negotiation Intelligence	29–40	Price evolution, outcomes, dialogue telemetry

DATASET SPECIFICATIONS

Dataset Scale

Metric	Value
Total Records	1,000,000
Total Negotiation Trajectories	125,000
Turns Per Trajectory	8
Features	40
Storage Format	Apache Parquet
Compression	Snappy
Recommended Memory	≥8 GB RAM

Core Research Characteristics

Structured Decision Intelligence

Each trajectory represents a sequential bargaining process containing:

Initial offer
Counter-offers
Concession progression
Strategic posture shifts
Resolution outcomes

This structure enables direct formulation of Markov Decision Processes (MDPs).

Multi-Agent Interaction Space

Each negotiation episode contains:

Distributor agent
Retail agent
Environmental state variables
Financial state variables
Action space
Outcome signals

This enables simulation of decentralized economic coordination systems.

Graph Intelligence Layer

Persistent identifiers allow construction of interaction graphs:

$A = (N, E)$

where nodes represent market participants and edges represent transactional relationships.

Potential tasks include:

Link prediction
Trust propagation
Community discovery
Credit network analysis
Graph representation learning

TELEMETRY ACQUISITION FRAMEWORK

Distributed Data Processing Architecture

KiranaChain is structured using a multi-layer telemetry architecture consisting of:

Layer 1 — Transaction State Capture

Captures:

Negotiation state transitions
Credit allocations
Outstanding liabilities
Settlement decisions

Layer 2 — Market Context Integration

Enriches observations with:

Commodity signals
Inflation indicators
Transportation conditions
Seasonal effects

Layer 3 — Conversational Intelligence Processing

Processes multilingual negotiation interactions into:

Structured dialogue logs
State transitions
Sentiment representations
Negotiation posture annotations

Layer 4 — Validation Engine

Performs:

Schema verification
Constraint enforcement
Cross-field consistency checks
Distribution validation

MATHEMATICAL INTEGRITY & STATISTICAL PROPERTIES

Non-Stationary Negotiation Dynamics

Negotiation outcomes evolve according to changing state conditions.

The probability of successful agreement can be expressed as:

$P (S_{t}) = f (T, L, D, E, t)$

where:

$$(T)$$ = trust
$$(L)$$ = liquidity
$$(D)$$ = demand conditions
$$(E)$$ = environmental variables
$$(t)$$ = negotiation stage

High-Dimensional Dependency Structure

The dataset contains complex relationships among:

Credit exposure
Liquidity stress
Trust scores
Inflation signals
Supply disruptions

These interactions create non-linear decision surfaces suitable for advanced machine learning systems.

Reinforcement Learning Compatibility

Each row can be interpreted as:

$(s_t,a_t,r_t,s_{t+1})$

where:

$$(s_t)$$ = negotiation state
$$(a_t)$$ = offer adjustment
$$(r_t)$$ = economic outcome
$$(s_{t+1})$$ = subsequent state

This supports:

Offline RL
Policy optimization
Multi-agent RL
Strategic planning systems

RESEARCH APPLICATIONS

Reinforcement Learning

Offline RL
Policy evaluation
Counterfactual reasoning
Negotiation optimization

Multi-Agent Systems

Cooperative bargaining
Competitive negotiation
Resource allocation
Distributed coordination

Financial Intelligence

Credit scoring
Risk estimation
Liquidity forecasting
Trust modeling

Supply-Chain Analytics

Disruption forecasting
Demand estimation
Inventory optimization
Logistics planning

Conversational AI

Dialogue modeling
Negotiation assistants
Strategy generation
Multilingual reasoning

DATA QUALITY ASSURANCE

Validation procedures include:

Structural Validation

Column integrity verification
Type consistency checks
Constraint enforcement

Statistical Validation

Distribution monitoring
Correlation verification
Outlier detection

Trajectory Validation

Sequential consistency
Episode completeness
Resolution verification

Conversational Validation

Dialogue schema compliance
Language consistency
Metadata integrity

ETHICAL CONSIDERATIONS

KiranaChain is intended exclusively for research, educational, benchmarking, and scientific evaluation purposes.

Researchers should carefully evaluate:

Generalization assumptions
Regional deployment considerations
Fairness implications
Economic decision impacts

The dataset should not be used as the sole basis for automated lending, credit approval, or financial exclusion decisions without rigorous domain-specific validation.

CITATION

@misc{atlas_ai_labs_2026,
    author       = { Atlas AI Labs and Dhadi Sai Praneeth Reddy and Mididuddi Dhatri and Biradar Amulya and Dr. Machana Jithender Reddy },
    title        = { KiranaChain (Revision 64b6d15) },
    year         = 2026,
    url          = { https://huggingface.co/datasets/Atlas-AI-Labs/KiranaChain },
    doi          = { 10.57967/hf/9015 },
    publisher    = { Hugging Face }
}

Downloads last month: 49

Total file size:

13.3 kB