Using the Export

Two practical ways to work with the exported data: building fine-tuning datasets and auditing compliance violations.

Supervised Fine-Tuning Dataset

Filter executions.jsonl for high-quality conversations and reformat them for your training pipeline:

Example

import json

with open("executions.jsonl") as f:
    executions = [json.loads(line) for line in f if line.strip()]

dataset = []
for e in executions:
    report = e.get("report") or {}
    if (
        report.get("is_completed") 
        and report.get("is_valid") 
        and report.get("is_factual") 
        and e.get("conversation")
    ):
        dataset.append({
            "messages": [
                {"role": m["role"], "content": m["content"]}
                 for m in e["conversation"]
            ]
        })

with open("sft_dataset.jsonl", "w") as out:
    out.writelines(json.dumps(row) + "\n" for row in dataset)

Offline Compliance Analysis

Combine the principles and report fields to identify which principles were violated and at what severity:

Example

import json

with open("executions.jsonl") as f:
    executions = [json.loads(line) for line in f if line.strip()]

for e in executions:
    report = e.get("report") or {}
    severity = report.get("compliance_violation_severity", 0)
    if severity > 0:
        print(f"Execution {e['id']}, severity {severity}")
        for p in e.get("principles") or []:
            print(f"  Principle: {p['name']}")