federalai/nifi-diagnosis-dev

NiFi Diagnosis Dev

Can an AI assistant explain a data flow without inventing what happened at runtime?

Twenty synthetic Apache NiFi configuration cases, explicit evidence-linked labels, and the complete raw responses from a deterministic baseline and a pinned local Qwen3-0.6B run. Prepared by Lloyd Clark, PhD · Federal.AI as part of Applied AI Flight Tests.

Interactive Space · Annotation rules · Source and grader · Federal.AI

What this tests

The narrow task is structured configuration diagnosis: identify four defined conditions, cite their JSON pointers, and decline unsupported runtime conclusions. Cases include nested process groups, legal zero timers, stopped processors, queue thresholds, missing processor endpoints, and questions that need telemetry.

One public dev split. These 20 hand-authored, AI-assisted fixtures and labels are suitable for prompt development, grader checks, and inspecting model failures. They are not an independent or held-out benchmark. The flows are simplified projections, not deployable NiFi exports.

Recorded result

Strict contract	Deterministic baseline	Qwen3-0.6B, recorded configuration
Schema-valid outputs	20 / 20	0 / 20
Exact findings, evidence and abstention	20 / 20	0 / 20

Every recorded model response contained Markdown fences, violating the raw-JSON contract; no repair was applied. Some also contain semantic errors. The baseline implements the same published rules as the labels, so its perfect score is a consistency check. The model result applies to this prompt and fixed greedy, non-thinking configuration, not Qwen's recommended sampling settings, larger models, or NiFi diagnosis generally. Full interpretation and runtime provenance.

Use it

Load the published dev split with the datasets library:

import json
from datasets import load_dataset

cases = load_dataset("federalai/nifi-diagnosis-dev", split="dev")
case = cases[0]
input_record = json.loads(case["input_json"])
expected_record = json.loads(case["expected_json"])
print(case["qwen_output"])  # Exact decoded response, including formatting failures.

Use the original prompt and grader for comparable evaluation. Keep labels and recorded outputs out of model inputs.

Provenance and limits

All 20 rows preserve the source input, label and prediction records as serialized strings. Raw model text is also exposed directly for inspection. Provenance pins the source commit and SHA-256 hashes; python verify_dataset.py reconstructs and verifies all four original JSONL files from the dataset columns.

Created for this exercise with AI assistance; no customer exports, personal information, private networks, or NiFi runtime telemetry were used. Runtime question type is explicit, so abstention measures compliance with a supplied boundary. The data cannot establish production reliability, calibrated uncertainty, disconnected deployment readiness, or generalization to new flows. No model weights are redistributed. Apache NiFi and Qwen remain their respective projects; this is an independent evaluation.

License: Apache-2.0. Maintainer: Lloyd Clark / Federal.AI. Cite this release using CITATION.cff.