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()]

def passed(report, dimension):
    """True only if every scored turn for this dimension passed."""
    decisions = [o["decision"] for o in report.get("outcomes", [])
                 if o["type"] == dimension and o["decision"] is not None]
    return bool(decisions) and all(decisions)

dataset = []
for e in executions:
    report = e.get("report") or {}
    if (
        passed(report, "completion")
        and passed(report, "validity")
        and passed(report, "factuality")
        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 {}
    # Worst compliance severity across all turns (0 = no violation).
    severity = max(
        (o.get("severity", 0) for o in report.get("outcomes", [])
         if o["type"] == "compliance"),
        default=0,
    )
    if severity > 0:
        print(f"Execution {e['id']}, severity {severity}")
        # Drill into per-principle assessments for the offending turns.
        for c in report.get("classifications", []):
            if c["type"] != "compliance":
                continue
            for a in c.get("assessments", []):
                if not a["decision"]:
                    print(f"  Principle {a['principle_id']}: {a['explanation']}")