IGS Dataset

AI Training Data Schema

IGS Dataset defines how interaction data is archived on Walrus for continuous AI model training, performance auditing, and protocol transparency.

Dataset Purpose

Every intent execution generates valuable training data:

• Intent Classification: Teach AI to categorize new intents

• Ranking Optimization: Learn which solutions perform best

• Solver Evaluation: Track historical performance patterns

• Market Analysis: Understand liquidity and pricing dynamics

Dataset Schema Reference

Schema Overview

Definition	Description
RawFeatures	Raw features extracted from IGSIntent for ML training
GroundTruthLabel	Ground truth classification labels
LabelingMetadata	Labeling metadata
ExecutionOutcome	Execution outcome for feedback loop
IntentClassificationTrainingData	Complete training sample for intent classification

RawFeatures

Property	Type	Description
solver_window_ms	number	Solver access window duration
user_decision_timeout_ms	number	User decision timeout
time_to_deadline_ms	number	Time until deadline
time_in_force	string	Time in force policy
max_slippage_bps	number	Maximum slippage in basis points
max_gas_cost_usd	number	Maximum gas cost in USD
max_hops	number	Maximum routing hops
has_whitelist	boolean	Has protocol whitelist
has_blacklist	boolean	Has protocol blacklist
has_limit_price	boolean	Has limit price condition
optimization_goal	string	Primary optimization goal
surplus_weight	number	Surplus weight (0-100)
gas_cost_weight	number	Gas cost weight (0-100)
execution_speed_weight	number	Execution speed weight (0-100)
reputation_weight	number	Reputation weight (0-100)
require_simulation	boolean	Requires simulation
input_count	number	Number of input assets
output_count	number	Number of output assets
input_asset_types	array	Input asset type classifications
output_asset_types	array	Output asset type classifications
input_value_usd	number	Input value in USD
expected_output_value_usd	number	Expected output value in USD
benchmark_source	string	Benchmark data source
benchmark_confidence	number	Benchmark confidence (0-1)
expected_gas_usd	number	Expected gas cost in USD
expected_slippage_bps	number	Expected slippage in basis points
has_nlp_input	boolean	Has natural language input
nlp_confidence	number	NLP parsing confidence (0-1)
client_platform	string	Client platform identifier
tag_count	number	Number of tags

Time in Force Values

Value	Description
immediate	Immediate execution
good_til_cancel	Good until canceled
fill_or_kill	Fill or kill order

Optimization Goal Values

Value	Description
maximize_output	Maximize output amount
minimize_gas	Minimize gas costs
fastest_execution	Fastest execution speed
balanced	Balanced optimization

Asset Type Values

Value	Description
native	Native blockchain token
stable	Stablecoin
volatile	Volatile cryptocurrency

GroundTruthLabel

Property	Type	Description
primary_category	string	Primary intent category
detected_priority	string	Detected user priority
complexity_level	string	Intent complexity level
risk_level	string	Associated risk level

Primary Category Values

Value	Description
swap	Token swap operation
limit_order	Limit order execution
complex_defi	Complex DeFi strategy
arbitrage	Arbitrage opportunity
other	Other intent type

Detected Priority Values

Value	Description
speed	Speed-focused
cost	Cost-focused
output	Output-focused
balanced	Balanced priority

Complexity Level Values

Value	Description
simple	Simple operation
moderate	Moderate complexity
complex	High complexity

Risk Level Values

Value	Description
low	Low risk
medium	Medium risk
high	High risk

LabelingMetadata

Property	Type	Description
labeling_method	string	Method used for labeling
labeled_by	string	Entity that created the label
labeled_at	number	Labeling timestamp
label_confidence	number	Label confidence (0-1)
notes	string	Additional labeling notes

Labeling Method Values

Value	Description
expert_manual	Manual expert labeling
rule_based	Rule-based automatic labeling
outcome_based	Based on execution outcomes
user_feedback	User feedback labeling
synthetic	Synthetically generated

ExecutionOutcome

Property	Type	Description
executed	boolean	Whether intent was executed
chosen_solution_rank	number	Rank of chosen solution
chosen_solution_id	string	ID of chosen solution
actual_metrics	object	Actual execution metrics
user_satisfaction	number	User satisfaction score (1-5)
executed_at	number	Execution timestamp

Actual Metrics Object

Property	Type	Description
actual_output_usd	number	Actual output value in USD
actual_gas_cost_usd	number	Actual gas cost in USD
actual_execution_time_ms	number	Actual execution time in milliseconds
actual_slippage_bps	number	Actual slippage in basis points

IntentClassificationTrainingData

Property	Type	Description
sample_id	string	Unique sample identifier
intent_metadata	object	Intent metadata object
raw_features	RawFeatures	Raw feature data
ground_truth	GroundTruthLabel	Ground truth labels
label_info	LabelingMetadata	Labeling metadata
execution_outcome	ExecutionOutcome	Execution outcome data
dataset_version	string	Dataset version identifier
created_at	number	Sample creation timestamp
full_intent_ref	string	Reference to full intent data

Intent Metadata Object

Property	Type	Description
intent_id	string	Intent identifier
intent_type	string	Intent type classification
created_at	number	Intent creation timestamp

Data Storage

Walrus Integration

All dataset entries stored on Walrus:

• Cost Efficiency: Cheap decentralized storage

• Availability: Distributed redundancy

• Integrity: Cryptographic verification

• Permanence: Long-term archival

Continuous Learning

Model Retraining Cycle

1. Data Collection: New executions archived hourly

2. Preprocessing: Anonymization and validation

3. Training: Updated models trained weekly

4. Validation: Performance tested on holdout set

5. Deployment: Gradual rollout to production

Reinforcement Learning

Positive Examples:

• Solutions that outperformed estimates

• High user satisfaction indicators

• Novel successful strategies

Negative Examples:

• Solutions that underperformed

• Failed executions

• User complaints or reverts

Learn More

• IGS Core - Data type definitions

• AI-Powered Ranking - How data is used

• Privacy Protection - Security details

References

6KB

dataset-schema.json

Open