shaunak1234/snmp-diagnostic-agent-traces

SNMP Diagnostic Agent Traces

2,136 tool-calling trajectories in which an agent diagnoses why SNMP discovery fails for a network device. It uses ping, traceroute, snmp_get, snmp_walk and read-only device show commands.

Every tool output comes from a deterministic simulator driven by a hidden network state. Every trajectory is replayed and checked against that state before it's accepted. No LLM wrote any tool output, root cause, or final answer.

Why this dataset exists

Free-form synthetic agent data has a known failure mode: the generating model invents plausible commands, outputs and conclusions, and the fine-tuned model learns to invent them too. This dataset inverts the process:

1. Sample a network state (poller, routed path, device, SNMP config, ACLs, supported OIDs) and inject one of 9 root causes.
2. A simulator renders realistic raw output: iputils ping, traceroute, net-snmp, Cisco IOS.
3. An oracle policy that sees only tool output picks the next tool until it can conclude.
4. The final answer is built only from observed facts.
5. Validators replay every call, check the label and required evidence, and check that every IP, number, OID and quoted string in the answer appears in the conversation.

Things the dataset gets right that most don't

• A wrong SNMPv2c community is a timeout, not an "authorization error." Agents silently drop unknown-community and ACL-denied requests. So wrong_community, snmp_acl_blocks_poller, snmp_agent_disabled and udp161_filtered_in_path show byte-identical SNMP symptoms, and the agent has to inspect the device to separate them.
• Counterfactual sets. Each base network is instantiated with all 9 causes, under the same user message. A model can't infer the cause from wording.
• Wrong user hypotheses. A third of the prompts include a suspicion ("the firewall change blocked UDP 161"). The agent must confirm or refute it with evidence.
• Honest failures. Where the oracle was misled (e.g. a lossy link that dropped 0 of 4 pings), the sample was rejected, not patched.

Root causes

Label	Distinguishing evidence
device_down	100% ping loss; traceroute reaches the device's /24 and stops
path_failure	100% loss; traceroute stops before the device's subnet, sometimes !N
packet_loss	ping loss ≥ 20%
snmp_agent_disabled	ping OK, SNMP timeout, no snmp-server community lines
wrong_community	ping OK, SNMP timeout, supplied community absent from config
snmp_acl_blocks_poller	community present; its ACL doesn't permit the poller
udp161_filtered_in_path	device config is correct and permits the poller; SNMP still times out
unsupported_oid	No Such Object for the discovery OID; the agent answers other OIDs
healthy	ping clean; the discovery OID returns data

Splits

split	samples	base networks	prompt templates
train	1,459	82	12
validation	197	11	12 (same as train)
test	480	27	6 held out

Splits are by base network, so every variant of a network stays in one split. The test split also uses prompt phrasings never seen in training.

Example (abridged)

user: Maybe the new firmware dropped support for ENTITY-MIB::entPhysicalSerialNum.1. Anyway,
      dist-sw-15.del2.netoai.lan is failing discovery - community snmp_ro, object ... (1.3.6.1.2.1.47.1.1.1.1.11.1).
→ ping dist-sw-15.del2.netoai.lan            5 received, 0% packet loss
→ snmp_get ... snmp_ro 1.3.6.1.2.1.47...     Timeout: No Response from dist-sw-15.del2.netoai.lan
→ show running-config | include snmp-server  (two community lines, neither is snmp_ro)
→ show snmp                                  3085 Unknown community name
assistant: The evidence does not point to the OID being unsupported. ... 'snmp_ro' is not one of them.
           SNMPv2c agents silently drop requests with an unknown community, which is why this looks like
           a timeout rather than an auth error. ...
           Diagnosis: wrong_community

Format

Each record is OpenAI-style chat with tool calls, ready for TRL / Axolotl / most chat templates:

{
  "id": "g0007-wrong_community-v1",
  "root_cause": "wrong_community",
  "difficulty": "hard",
  "user_hypothesis": "oid",
  "template_id": "hard_6",
  "tool_sequence": ["ping", "snmp_get", "device:show running-config | include snmp-server", "device:show snmp"],
  "messages": [
    {"role": "system", "content": "..."},
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "Start with basic reachability...", "tool_calls": [{"id": "call_…", "type": "function", "function": {"name": "ping", "arguments": "{\"host\": \"…\", \"count\": 5}"}}]},
    {"role": "tool", "tool_call_id": "call_…", "name": "ping", "content": "PING …"},
    {"role": "assistant", "content": "… \n\nDiagnosis: wrong_community"}
  ],
  "tools": [ … 5 function schemas … ]
}

ground_truth/ holds the hidden network states (states.jsonl), the canonical traces (trajectories.jsonl), the rejected samples with reasons (rejected.jsonl), and generation stats.

Benchmark

The test split doubles as an interactive benchmark. The agent calls the simulator live, and the harness scores:

• accuracy
• correct + sufficient evidence: the right label and the observations that justify it
• invalid tool-call rate, ungrounded claims, and community-string leaks

Agent	Accuracy	Correct + evidence	Community-string leaks
Rule-based oracle	99.6%	99.6%	0.0%
Qwen3-8B + LoRA on the train split	85.4%	84.6%	1.7%
ping + snmp heuristic	53.8%	20.0%	0.0%
Qwen3-8B, prompted (no fine-tune)	50.4%	38.3%	8.8%
One snmp_get, then guess	33.8%	0.0%	0.0%

The LoRA run: rank 16 on all attention and MLP projections, 2 epochs, loss on assistant turns only, served with vLLM (enable_thinking=False). The hardest cause for the fine-tuned model is udp161_filtered_in_path (46%), which it mostly confuses with an SNMP ACL, a disabled agent or an unsupported OID.

If you fine-tune Qwen3 with thinking disabled, render every assistant turn with the empty <think>\n\n</think>\n\n block that the generation prompt adds at inference. The stock chat template only puts it before the last turn. Training without it teaches the model that the empty block means "answer now", and it stops after about one tool call. In our first run, that gave 46% instead of 85%.

The generator, simulator, validators and harness are in the companion code repository.

Limitations

• Simulated, not captured, output. The formats follow real tools, but device behavior is limited to what the simulator models (Cisco IOS-style SNMPv2c, standard ACLs).
• Single-fault scenarios. Real incidents can combine causes.
• Templated language: 18 user templates and 3 system prompts. The companion code includes an optional, re-validated LLM paraphrase step.
• udp161_filtered_in_path is concluded by eliminating device-side causes, which is how an engineer would reason, but a packet capture would be the definitive proof.

License

Apache 2.0