hbnunlp/NetConfigQA2.1

NetConfigQA 2.1

NetConfigQA 2.1 is a benchmark for evaluating whether LLMs can infer network behavior from raw router configuration files (Cisco IOS .cfg). It contains 9,473 QA instances across four emulated SP/DC topologies (10–40 nodes) and five difficulty levels, from single-device fact extraction (L1) to what-if failure analysis (L5). All ground-truth answers are generated automatically from the configurations using Batfish (static analysis + simulation). All 5,719 L1–L3 answers were cross-checked against an independent, Batfish-free regex parser (99.91% agreement).

Repository layout

data/LabA.jsonl ... LabD.jsonl     # QA instances, one JSON object per line
labs/LabA/configs/*.cfg            # device configurations (model input)
labs/LabB/configs/*.cfg
labs/LabC/configs/*.cfg
labs/LabD/configs/*.cfg

Topologies and QA distribution

Lab	Nodes	Concept	Key protocols	L1	L2	L3	L4	L5	Total	QA/N
Lab-A	10	SP MPLS VPN	OSPF, BGP, LDP, VRF	694	70	252	153	103	1,272	127
Lab-B	20	Extended SP	+ NTP, SNMP, AAA	1,274	57	255	443	128	2,157	108
Lab-C	30	Security + L2VPN	+ L2VPN, ACL, eBGP, HSRP	1,271	39	255	954	155	2,674	89
Lab-D	40	Multi-AS Complex	+ QoS, NetFlow, 3-AS	1,271	28	253	1,657	161	3,370	84
Total				4,510	194	1,015	3,207	547	9,473

Difficulty levels

Level	Domain	Reasoning	Example
L1	Static	Fact extraction from a single device	What is the system hostname on p4?
L2	Static	Aggregation across all devices	Which devices have SSH enabled?
L3	Static	Cross-device consistency comparison	Which L2VPN pairs have only unidirectional configuration?
L4	Dynamic	Packet-path simulation (traceroute / reachability)	What is the traceroute path from 172.16.4.2 to 10.0.0.5?
L5	Dynamic	What-if analysis under link/device failure	Which devices are single points of failure (SPOF)?

Data fields

Field	Description
uid	Globally unique ID (<lab>/<id>)
lab	LabA – LabD
id	Question ID, unique within a lab
metric	Metric (question template) name
category	Metric category (e.g. System_Inventory, OSPF_Consistency, Reachability_Analysis)
level	Difficulty level L1–L5
scope	JSON string describing the question target (DEVICE, GLOBAL, NODE_PAIR, ...)
question	Question text, including the expected answer format
answer_type	text, number, set_str, set, map_str_str, map_str_int
answer	Raw ground-truth answer (JSON-encoded string)
eval_answer	Normalized answer used for scoring
oracle_source	How the answer was derived: parser (Batfish config parsing), hybrid (parsed facts + cross-device logic), batfish_replay (Batfish simulation, L4/L5)
files	Configuration files (under labs/<lab>/configs/) needed to answer. An empty list means the whole lab (all L4/L5 questions require the full topology)

Evaluation: Type-Aware Accuracy (TA-Acc)

Embedding-based similarity (e.g. BERTScore) does not separate correct from incorrect answers on network data, so each answer type uses its own comparison function:

Answer type	Comparison	Example
Set (set, set_str)	Set F1 (order-agnostic)	BGP neighbor list, interface set
Map (map_str_str, map_str_int)	Key-value F1	Per-device attributes / statistics
Number (number)	Numeric normalization + exact match	Hop count, AS number
Text (text)	Normalized exact match (order-preserving)	Hostname, IP address, traceroute path

NOT_CONFIGURED answers

2,082 instances (22.0%) ask about a feature that is not configured on the target device(s), e.g. a timezone on a router without clock timezone. For these, eval_answer is NOT_CONFIGURED. A model is expected to recognize the absence of the feature rather than hallucinate a value.

Notes

• Sanitization. Local-user secret values in the configurations were replaced with a dummy hash. No question depends on them.
• Known limitation. 8 L5 config-change questions ("... after the configuration modification?") do not state which device was modified in the question text.