netcfg-config-json

2,284 verified (network configuration to normalized JSON) pairs across 17 vendor platforms, plus the evaluation harness used to measure them.

Built with netlab (MIT). Every configuration is real, vendor-valid syntax emitted by netlab's own device templates from randomly generated topologies.

Türkçe özet: 17 farklı üretici platformu için 2.284 adet doğrulanmış konfigürasyon ve normalize JSON çifti, ve bunları ölçmek için kullanılan değerlendirme kodu.

Why this exists

Organizations cannot send device configurations and incident tickets to external AI services. Configs contain topology, addressing, and policy — the map of the network. A model that runs entirely on local hardware removes that objection. But building one requires a dataset, and there is no public corpus of configuration-to-structure pairs. This is an attempt at one.

Contents

File	Rows	Description
dataset.jsonl	2,284	raw pairs: id, device, config, label
sft/train.jsonl	2,541	ChatML, training split
sft/val.jsonl	286	ChatML, held-out split

Two tasks are present. parse (config to JSON) is the primary one: 2,048 train / 236 validation. render (JSON to config) is secondary: 493 / 50.

Device coverage

Platform	n	Platform	n
routeros7	400	dellos10	73
eos	367	cat8000v	50
nxos	268	csr	45
frr	210	ioll2	36
srlinux	161	iol	28
fortios	157	iosvl2	23
sros	129	sonic	3
iosxr	129
arcos	115
arubacx	90

Coverage is deliberately uneven toward platforms with narrow module support (FortiOS supports only BGP, OSPF, and routing in netlab), so they are not crowded out by platforms supporting 15+ modules.

Two design decisions worth knowing about

1. Every label is derivable from the config text

The generator has access to netlab's full topology model, which knows far more than any single device's configuration reveals — peer hostnames, link metadata, global intent. Training a model to output facts it cannot see in its input teaches it to hallucinate.

So labels are filtered to fields that provably appear in the config text. Interface labels keep only ifname, ipv4, ipv6, mtu, type, vrf, vlan, bandwidth. BGP neighbor labels keep as, ipv4, ipv6, type, local_as, rr — and drop name, because the neighbor's hostname is not in the configuration.

A verification pass then measures, for every sample, the fraction of label facts findable in the config string. All 2,284 samples score 1.000.

2. The split is by topology, not by sample

Configurations from the same topology share addressing, AS numbers, and interface naming. Splitting randomly at the sample level would place a router's neighbors in training and the router itself in validation — leakage that inflates scores.

The split is applied at the topology seed level. No validation configuration comes from a topology that appears in training.

Evaluation harness

The scoring code is included:

• scoring.py — field-level F1 with order-independent flattening (a JSON object with the same interfaces in a different order scores identically), JSON validity, exact match, and a hallucination rate computed only over valid-JSON rows (counting invalid JSON as "clean" would reward failure).
• run_eval.py — runs any OpenAI-compatible endpoint (LM Studio, llama.cpp server, vLLM) and writes a comparison report.
• schema_prompt.py — the schema description given to baseline models, so an untrained model is measured on a task it has been told how to do.
• build_regression.py / run_regression.py — a 200-question MMLU subset for catastrophic-forgetting checks.

Reference results

Qwen3-8B, Q4_K_M, 236 held-out configurations, temperature=0:

Setup	Field F1	JSON validity	Hallucination
Base, no schema in prompt	0.0227	0.8686	0.0008
Base, schema in prompt	0.5752	0.9492	0.0104
Fine-tuned (LoRA, 2 epochs)	0.9219	1.0000	0.0073

The first row is worth dwelling on: without being told the target shape, the model scores essentially zero. Any published before/after comparison that uses a zero-shot baseline is measuring prompt engineering, not fine-tuning.

Trained model: huseyinturkdogan/netcfg-8b-lora · huseyinturkdogan/netcfg-8b-GGUF

Limitations

This is synthetic data. Vendor-valid syntax, generated from real device templates, but not production configuration. Real configs carry comments, disabled blocks, historical cruft, site conventions, and vendor quirks absent here. A model trained only on this will degrade on production text.

Module coverage follows netlab's, not the vendors'. Features netlab does not model do not appear.

SONiC has 3 samples. Treat anything below ~50 samples as anecdotal.

Attribution

Configurations generated with netlab by Ivan Pepelnjak and contributors (MIT). General instruction data used in the training mix comes from mlabonne/FineTome-100k.