example_id stringlengths 18 59 | kind stringclasses 2
values | profile dict | intent stringlengths 27 81 | before unknown | trajectory listlengths 2 6 | reply stringlengths 62 225 | pipeline unknown | prompt_version stringclasses 1
value |
|---|---|---|---|---|---|---|---|---|
churn-risk-telecom | create | {
"industry": "telecom",
"persona": {
"role": "data analyst",
"constraints": [
"limited computational resources"
]
},
"connections": [
{
"name": "crm.customers",
"state": "captured"
}
],
"tables": [
{
"database": "crm",
"name": "customers",
"rows":... | Predict customer churn risk for proactive retention strategies | null | [
{
"call": "list_datasets",
"args": {},
"observation": "['crm.customers']"
},
{
"call": "get_dataset",
"args": {
"dataset": "crm.customers"
},
"observation": "Columns: customer_id, subscription_status, last_contact_date, total_calls, total_complaints, monthly_fee, churn_label; l... | I've set up a predict pipeline for customer churn risk prediction. I've removed customers with no last_contact_date to improve time-based analysis. The model will predict churn based on available features. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
predict-employee-departure | create | {
"industry": "human_resources",
"persona": {
"role": "HR analyst",
"constraints": [
"needs predictive insights"
]
},
"connections": [
{
"name": "hr.employee_data",
"state": "captured"
}
],
"tables": [
{
"database": "hr_db",
"name": "employee_data",
... | Predict which employees are likely to leave the company | null | [
{
"call": "list_datasets",
"args": {},
"observation": "found dataset hr.employee_data"
},
{
"call": "get_dataset",
"args": {
"dataset": "hr.employee_data"
},
"observation": "resignation_notice_date (10% non-null), event_updated_datetime_pt (no nulls)",
"because": "list_data... | I've built a predictive model to identify employees likely to leave based on their promotion history and other factors. The model is calibrated and validated with a 20% holdout. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
churn-calibration-mutation | mutate | {
"industry": "telecommunications",
"persona": {
"role": "data scientist",
"constraints": [
"tight deadline"
]
},
"connections": [
{
"name": "accounts",
"state": "captured"
}
],
"tables": [
{
"database": "telecom",
"name": "accounts",
"rows": 90000... | add calibration to the existing churn prediction model | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": false,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "telecom.accounts returns (4 cols, 900K rows)"
},
{
"call": "get_dataset",
"args": {
"dataset": "telecom.accounts"
},
"observation": "customer_id (0% null), churn_flag (5% null), signup_date, last_interaction (timestamp)",
... | I've added probability calibration to the churn prediction model to improve the accuracy of probability scores. The updated model is ready to be deployed. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
lead-scoring-timeframe-adjustment | mutate | {
"industry": "sales",
"persona": {
"role": "sales strategy analyst",
"constraints": [
"needs recent data for decision making"
]
},
"connections": [
{
"name": "crm",
"state": "captured"
}
],
"tables": [
{
"database": "crm",
"name": "sales_leads",
"... | change the model to use only data from the past 6 months | {
"connect": true,
"clean": true,
"point_in_time": false,
"leakage_free": false,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "crm.sales_leads (4 cols, 500k rows)",
"because": "first call—initial exploration to understand existing data"
},
{
"call": "get_dataset",
"args": {
"dataset": "crm.sales_leads"
},
"observation": "lead_id and interaction... | The model has been updated to only consider data from the past 6 months. This helps in meeting the need for recent data in your decision-making process. | null | v1 |
retail-forecast-add-backtest | mutate | {
"industry": "retail",
"persona": {
"role": "data scientist",
"constraints": [
"build a demand forecasting model"
]
},
"connections": [
{
"name": "forecasting_database",
"state": "captured"
}
],
"tables": [
{
"database": "forecasting_database",
"name": ... | Add a backtest to assess forecasting stability over time. | {
"connect": true,
"clean": true,
"temporal_validation": true,
"holdout": true,
"model": true,
"model_kind": "forecast",
"serve": true,
"act": true
} | [
{
"call": "get_dataset",
"args": {
"dataset": "forecasting_database.sales_data"
},
"observation": "columns include date, product_id, units_sold, store_id",
"because": "sales_data was used in the existing pipeline and still appears suitable"
},
{
"call": "model",
"args": {
... | I've added a backtest to your forecasting pipeline to assess its stability across different time periods. | {
"connect": true,
"clean": true,
"temporal_validation": true,
"backtest": true,
"holdout": true,
"model": true,
"model_kind": "forecast",
"serve": true,
"act": true
} | v1 |
loan-default-classification | create | {
"industry": "finance",
"persona": {
"role": "data scientist",
"constraints": [
"ensure model calibration",
"avoid data leakage"
]
},
"connections": [
{
"name": "loans_db",
"state": "captured"
}
],
"tables": [
{
"database": "loans_db",
"name": "lo... | predict which loans are likely to default | null | [
{
"call": "list_datasets",
"args": {},
"observation": "loans_db.loans (5 cols, 500k rows)"
},
{
"call": "get_dataset",
"args": {
"dataset": "loans_db.loans"
},
"observation": "loan_id, customer_id, loan_amount, issue_date, default_status (5% null)",
"because": "list_dataset... | The loan default prediction model is now built using a gradient boosting decision tree method, with calibration and a holdout set to ensure model reliability. The model can be used to classify the likelihood of loan defaults. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
ecommerce-churn-drop-null-accounts | mutate | {
"industry": "ecommerce",
"persona": {
"role": "data scientist",
"constraints": [
"ensure data quality"
]
},
"connections": [
{
"name": "ecommerce_db",
"state": "captured"
}
],
"tables": [
{
"database": "ecommerce_db",
"name": "customers",
"rows":... | drop null customer IDs in the churn model | {
"connect": true,
"clean": false,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "classifier",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "discovered ecommerce_db.customers table",
"because": "(first call — no justification needed)"
},
{
"call": "get_dataset",
"args": {
"dataset": "ecommerce_db.customers"
},
"observation": "10% null in the customer_id colu... | The null customer IDs have been dropped to improve data quality in the churn prediction model. Ensure all subsequent stages reflect these changes. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "classifier",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
healthcare-predict-readmissions-adjust-holdout | mutate | {
"industry": "healthcare",
"persona": {
"role": "data scientist",
"constraints": [
"models should have sufficient validation"
]
},
"connections": [
{
"name": "hospital_data",
"state": "captured"
}
],
"tables": [
{
"database": "hospital_data",
"name": "p... | Adjust the holdout size for the readmission prediction model | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": 0.2,
"explain": "shap",
"serve": true,
"act": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "hospital_data.patient_readmissions (30 cols, 500K rows)",
"because": "exploring data sources for existing pipeline adjustment"
},
{
"call": "get_dataset",
"args": {
"dataset": "hospital_data.patient_readmissions"
},
"ob... | The holdout size for the readmission prediction model has been adjusted to 30% for better validation. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": 0.30000000000000004,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
bank-fraud-detection-update-leakage | mutate | {
"industry": "banking",
"persona": {
"role": "Fraud Analyst",
"constraints": [
"real-time detection"
]
},
"connections": [
{
"name": "fraud.transactions",
"state": "captured"
}
],
"tables": [
{
"database": "frauddb",
"name": "transactions",
"rows"... | update the pipeline to consider potential leakage concerns | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": false,
"model": true,
"model_kind": "classifier",
"calibrated": true,
"holdout": 0.2,
"explain": "shap",
"serve": true,
"act": true
} | [
{
"call": "get_dataset",
"args": {
"dataset": "fraud.transactions"
},
"observation": "… approval_date can be a leakage source with transaction_date …",
"because": "the intent is to address potential leakage, so understanding table structure is needed"
},
{
"call": "clean",
"arg... | I've updated the pipeline to ensure it's leakage-free by removing approval_date, which could leak future information. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "classifier",
"calibrated": true,
"holdout": 0.2,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
car-insurance-claims-anomaly-detection | create | {
"industry": "insurance",
"persona": {
"role": "fraud investigator",
"constraints": [
"needs an unsupervised model for anomaly detection"
]
},
"connections": [
{
"name": "insurance.claims_data",
"state": "captured"
}
],
"tables": [
{
"database": "insurance",
... | Find anomalies in car insurance claims | null | [
{
"call": "list_datasets",
"args": {},
"observation": "insurance.claims_data"
},
{
"call": "get_dataset",
"args": {
"dataset": "insurance.claims_data"
},
"observation": "claim_id, claim_amount, customer_age, incident_type, report_date, incident_date",
"because": "list_datas... | Anomaly detection on car insurance claims is set up and ready to identify potential fraudulent claims. | {
"connect": true,
"clean": false,
"point_in_time": false,
"leakage_free": false,
"model": true,
"model_kind": "anomaly",
"threshold_governed": true,
"serve": true,
"act": false
} | v1 |
music-streaming-session-recommendation | create | {
"industry": "music streaming",
"persona": {
"role": "data scientist",
"constraints": [
"real-time predictions",
"large user base"
]
},
"connections": [
{
"name": "streaming.sessions",
"state": "captured"
},
{
"name": "content.songs",
"state": "captur... | build a recommendation system for suggesting songs during a session | null | [
{
"call": "list_datasets",
"args": {},
"observation": "streaming.sessions, content.songs"
},
{
"call": "get_dataset",
"args": {
"dataset": "streaming.sessions"
},
"observation": "session dataset with session_id, user_id, song_id, and timestamp",
"because": "list_datasets id... | Built a real-time song recommendation system based on past sessions and song metadata. Dropped sessions with missing play duration. | {
"connect": true,
"clean": true,
"recommendation": true,
"index": true,
"grounded_in_data": true
} | v1 |
supply-chain-predictive-maintenance-add-calibration | mutate | {
"industry": "manufacturing",
"persona": {
"role": "supply chain manager",
"constraints": [
"focus on preventive maintenance"
]
},
"connections": [
{
"name": "factory.machines",
"state": "captured"
}
],
"tables": [
{
"database": "factory",
"name": "mach... | add calibration to the existing predictive maintenance model | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "classifier",
"calibrated": false,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "factory.machines dataset available",
"because": "assess the available datasets for additional data if necessary"
},
{
"call": "get_dataset",
"args": {
"dataset": "factory.machines"
},
"observation": "failure_date (15% n... | The predictive maintenance model is now calibrated to improve reliability of failure predictions. Remaining steps like serving or acting are unchanged. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "classifier",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
adtech-click-fraud-detection | create | {
"industry": "adtech",
"persona": {
"role": "data scientist",
"constraints": [
"focus on fraud detection"
]
},
"connections": [
{
"name": "ad_clicks",
"state": "captured"
}
],
"tables": [
{
"database": "adtech",
"name": "ad_clicks",
"rows": 100000... | Detect fraudulent ad clicks | null | [
{
"call": "list_datasets",
"args": {},
"observation": "adtech.ad_clicks (6 cols, 1M rows)",
"because": ""
},
{
"call": "get_dataset",
"args": {
"dataset": "adtech.ad_clicks"
},
"observation": "is_fraud (2% non-null), ip_address (1% null), user_agent (5% null)",
"because... | I've created an anomaly detection pipeline to detect click fraud. Based on available data, I've cleaned nulls and set a classifier model to determine fraudulent clicks from labeled data. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "classifier",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
manufacturing-defective-products-detection-update-threshold | mutate | {
"industry": "manufacturing",
"persona": {
"role": "quality control engineer",
"constraints": [
"strict budget",
"limited computing resources"
]
},
"connections": [
{
"name": "iot.sensors",
"state": "captured"
},
{
"name": "crm.orders",
"state": "disc... | adjust the anomaly detection threshold for detecting defective products | {
"connect": true,
"clean": true,
"model": true,
"model_kind": "anomaly",
"threshold_governed": true,
"threshold_value": 0.75,
"serve": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "iot.sensors (4 cols, 1M rows)",
"because": "first call — no justification needed"
},
{
"call": "get_dataset",
"args": {
"dataset": "iot.sensors"
},
"observation": "anomaly_score available with no nulls; suitable for thr... | The threshold for detecting defective products has been adjusted to 0.65 to better balance false positives and false negatives within your resource constraints. | {
"connect": true,
"clean": true,
"model": true,
"model_kind": "anomaly",
"threshold_governed": true,
"threshold_value": 0.65,
"serve": true
} | v1 |
telecom-churn-modify-calibration | mutate | {
"industry": "telecommunications",
"persona": {
"role": "data scientist",
"constraints": [
"calibration needs improvement"
]
},
"connections": [
{
"name": "customer-churn",
"state": "captured"
}
],
"tables": [
{
"database": "telecom",
"name": "customer_... | improve calibration on the existing churn model | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": false,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "customer-churn dataset named",
"because": "(first call — no justification needed)"
},
{
"call": "get_dataset",
"args": {
"dataset": "telecom.customer_data"
},
"observation": "last_login has 20% null values; churn column... | I have improved the churn model's calibration by enabling it this time, which will now give more reliable probabilities of customer churn. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
hr-employee-turnover-adjust-holdout | mutate | {
"industry": "human_resources",
"persona": {
"role": "HR Analyst",
"constraints": [
"must maintain data confidentiality"
]
},
"connections": [
{
"name": "hr.employee_data",
"state": "captured"
}
],
"tables": [
{
"database": "hr",
"name": "employee_data"... | increase the holdout size for more robust validation | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": 0.2,
"explain": "shap",
"serve": true,
"act": true
} | [
{
"call": "list_pipelines",
"args": {},
"observation": "hr-employee-turnover-pipeline exists with 20% holdout"
},
{
"call": "update_pipeline",
"args": {
"holdout": 0.30000000000000004
},
"observation": "holdout updated from 20% to 30%",
"because": "list_pipelines showed tha... | I have successfully increased the holdout size to 30% for your employee turnover prediction model. This will enhance the robustness of your model's validation. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": 0.30000000000000004,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
energy-demand-forecasting | create | {
"industry": "energy",
"persona": {
"role": "data scientist",
"constraints": [
"seasonality aware",
"actionable insights needed"
]
},
"connections": [
{
"name": "weather.api",
"state": "captured"
}
],
"tables": [
{
"database": "energy",
"name": "h... | forecast the energy demand for the next month | null | [
{
"call": "list_datasets",
"args": {},
"observation": "energy.hourly_consumption dataset has hourly timestamps, consumption data, and temperature readings"
},
{
"call": "get_dataset",
"args": {
"dataset": "energy.hourly_consumption"
},
"observation": "columns: timestamp, consum... | I've built a forecasting model that takes into account hourly seasonality and the available temperature data to predict energy demand for the next month. The model includes a backtest for validation. | {
"connect": true,
"clean": true,
"temporal_validation": true,
"backtest": true,
"model": true,
"model_kind": "forecast",
"holdout": true,
"serve": true,
"act": true
} | v1 |
finance-personalized-credit-card-offers-change-model-kind | mutate | {
"industry": "finance",
"persona": {
"role": "marketing analyst",
"constraints": [
"optimize customer engagement"
]
},
"connections": [
{
"name": "banking.transactions",
"state": "captured"
}
],
"tables": [
{
"database": "banking",
"name": "transactions... | Can you change the model to focus on personalized credit card offers instead? | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "uplift",
"incrementality": true,
"holdout": true,
"serve": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "banking.transactions"
},
{
"call": "get_dataset",
"args": {
"dataset": "banking.transactions"
},
"observation": "columns: transaction_id, customer_id, amount, transaction_type, timestamp",
"because": "list_datasets name... | The model has been updated to focus on personalized credit card offers using a retrieval model centered around customer interactions. | {
"connect": true,
"clean": true,
"model": true,
"model_kind": "retrieval",
"index": true,
"grounded_in_data": true,
"serve": true
} | v1 |
healthcare-predict-readmissions-add-explanations | mutate | {
"industry": "healthcare",
"persona": {
"role": "data scientist",
"constraints": [
"explainability is crucial"
]
},
"connections": [
{
"name": "hospital_patient_records",
"state": "captured"
}
],
"tables": [
{
"database": "hospital_db",
"name": "admissi... | Explain predictions on the readmission model | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": true,
"serve": true,
"act": true
} | [
{
"call": "list_explainers",
"args": {},
"observation": "shap values, lime available",
"because": "the current pipeline lacks explanations and the user asked for them"
},
{
"call": "add_explanation",
"args": {
"method": "shap"
},
"because": "list_explainers showed SHAP avai... | I've added SHAP-based explanations to your readmissions model. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | v1 |
sports-analytics-match-prediction-update-model-kind | mutate | {
"industry": "sports analytics",
"persona": {
"role": "data scientist",
"constraints": [
"project deadline",
"limited resources"
]
},
"connections": [
{
"name": "sports_db",
"state": "captured"
}
],
"tables": [
{
"database": "sports_db",
"name": "... | switch from a classification model to a uplift model for enhanced decision making | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "classifier",
"calibrated": true,
"holdout": true,
"explain": "shap",
"serve": true,
"act": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "sports_db.match_data has columns match_id, team_id, match_date, result, elo_rating"
},
{
"call": "get_dataset",
"args": {
"dataset": "sports_db.match_data"
},
"observation": "result is fully populated, elo_rating has 5% nul... | The model kind has been switched to uplift for enhanced decision making around matches, using the 'result' column as the label. | null | v1 |
ecommerce-churn-change-model-kind | mutate | {
"industry": "ecommerce",
"persona": {
"role": "data scientist",
"constraints": [
"tight deadline"
]
},
"connections": [
{
"name": "crm_db",
"state": "captured"
}
],
"tables": [
{
"database": "crm_db",
"name": "user_behavior",
"rows": 150000,
... | switch the model to a classifier instead of the current gradient boosting model | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "gbdt",
"calibrated": true,
"holdout": true
} | [
{
"call": "list_datasets",
"args": {},
"observation": "crm_db contains user_behavior with 150k rows",
"because": "the previous pipeline used the crm_db user_behavior table"
},
{
"call": "get_dataset",
"args": {
"dataset": "crm_db.user_behavior"
},
"observation": "user_behav... | I've adjusted the model to use a classifier instead of a gradient boosting model. I've also reduced the holdout to accommodate recalibration on this model. The new setup is ready. | {
"connect": true,
"clean": true,
"point_in_time": true,
"leakage_free": true,
"model": true,
"model_kind": "classifier",
"calibrated": true,
"holdout": true
} | v1 |
telecom-network-downtime-forecasting | create | {
"industry": "telecom",
"persona": {
"role": "network manager",
"constraints": [
"minimize downtime"
]
},
"connections": [
{
"name": "network_outages",
"state": "captured"
},
{
"name": "network_traffic",
"state": "captured"
}
],
"tables": [
{
... | forecast network downtime to prevent outages | null | [
{
"call": "list_datasets",
"args": {},
"observation": "network_outages, network_traffic"
},
{
"call": "get_dataset",
"args": {
"dataset": "network_outages"
},
"observation": "outage_start_time and outage_duration indicate time intervals",
"because": "list_datasets named net... | Network downtime forecasting set up with temporal validation and backtesting due to timeseries data. | {
"connect": true,
"temporal_validation": true,
"backtest": true,
"holdout": true,
"model": true,
"model_kind": "forecast",
"serve": true,
"act": true
} | v1 |
project-agent-sft-corpus
SFT training data for the main agent — the one that creates and mutates Pipelines.
Private. Internal-confidential. Generated by zetta-bridge/modules/agent-project-corpus
(zetta#986).
** 22 examples · 14 mutations.**
It is a tool-calling agent, so these are trajectories
ask → call → observation → next call decided BY that observation → Pipeline
Every call after the first carries a because naming a previous observation. A call
justified by the user's ask is a scripted step, and a scripted trajectory is a lookup with extra
steps. That check rejects most first attempts, which is the point.
Quality bar: perturb one fact in the profile and the correct trajectory must change. If it survives, the example teaches emission from the use-case name rather than reading of the data.
Mutations are half the corpus, deliberately
Changing a Pipeline is the harder half of the product and the easier half to under-represent: "add calibration to that", "actually use the last 90 days", "that ran too long". A mutation carries the pipeline before — what must not change is as much of the lesson as what must, and one that silently rebuilds has lost the user's work.
SFT now, RLVR after — and the order is measured
RLVR sharpens what a policy already sometimes does; it cannot originate a capability. A campaign against a policy that never emits a pipeline returns a uniformly zero reward, and a group with equal rewards has zero advantage — GRPO takes a null step and cannot learn however long it runs. Measured, not assumed (zetta#859).
So this corpus has one job: get the policy to sometimes build a correct pipeline. The loop takes it from there — reward rose 0.52 → 1.00 over 16 steps against the deterministic graders.
Limitations
- ** 22 examples.** A first batch, not a corpus to train on.
- Roughly half of generations are rejected, now mostly for methodology-field errors rather than structural ones. Genuine rejections, but it means the kept examples skew toward whatever the model finds easy to get right.
- Nothing has been trained on this, so no evidence it teaches what it intends to.
- Synthetic, and the observations are written rather than returned by a real node — a trajectory here is what a good one looks like, not one that was actually executed.
- Downloads last month
- 6