KiranaChain
Atlas AI Labs
1M-turn multi-agent negotiation dataset for India's kirana economy: informal Udhar credit, supply-chain signals and multilingual bargaining dialogue.
- Negotiation-turn records
- 1,000,000
- Complete trajectories
- 125,000
- 8 turns each
- Features
- 40
01
Problem
- Who
- Researchers in reinforcement learning, multi-agent systems, credit risk and multilingual NLP.
- Why it mattered
- India's 12–14 million kirana stores run on negotiation, informal Udhar credit and code-switched bargaining, yet none of it exists in public structured datasets.
02
Architecture
- Component
- Key decision (click)
Agents
- Agent instantiation12 Indian states
- Financial + environmental copulasGaussian, Cholesky-decomposed
Negotiate
- Gaussian-process noiseSquared-exponential kernel, l = 5.0
Dialogue
- Posture assignment
- Hinglish / Telugu-English dialoguewith sentiment scoring
Ship
- Parquet outputSnappy, ~156 MB
- Validation & QAShape, nulls, leakage, correlations
A 12-stage procedural pipeline generates each 8-turn episode: copulas set correlated financial and environmental conditions, a non-stationary Markov chain resolves settlement, walk-away or credit denial, and GP noise adds human irrationality to price paths.
03
Key decisions
D1Non-stationary Markov chain with absorbing terminal states
- Alternatives considered
- Independent random outcomes
- Why
- Settlement, walk-away and credit denial depend on the turn and on liquidity stress, as in real bargaining.
04
Build highlights
- 01
Realistic dependencies
Trust score ↔ default risk correlates at r = −0.8429 (design target ≈ −0.78); liquidity stress ↔ default risk at +0.85.
- 02
Leak-free splits
80/10/10 train, validation and test split at trajectory level, with zero trajectory_id overlap verified.
- 03
Ready for many fields
Offline RL, settlement prediction, Udhar default modelling, code-switched dialogue, supply-chain forecasting and graph learning.
05
Results
61.23%
Settlement rate
20.47%
Walk-away rate
−0.8429
Trust ↔ default risk (r)
~128 s
Generation time
Colab CPU
How this was measured: Validation suite: dataset shape, null values, trajectory integrity, schema, correlation structure, split leakage and distribution consistency, all passed.
06
