---
license: mit
pretty_name: NetOps intent to CLI (synthetic operator-style requests)
language: [en]
tags: [networking, network-automation, cli, synthetic, instruction-tuning]
size_categories: [1K<n<10K]
task_categories: [text-generation]
configs:
  - config_name: default
    data_files:
      - {split: train, path: data/train.jsonl}
      - {split: validation, path: data/val.jsonl}
      - {split: test, path: data/test.jsonl}
  - config_name: v2
    data_files:
      - {split: train, path: v2/train.jsonl}
      - {split: validation, path: v2/val.jsonl}
      - {split: test, path: v2/test.jsonl}
      - {split: operator_test, path: v2/operator_test.jsonl}
---

# netops-intent2cli

**Synthetic operator-style English intents paired with verbatim CLI commands** from the 70,006-row [multivendor-cli-commands](https://huggingface.co/datasets/gesh75/multivendor-cli-commands) corpus.

Two configs: **default (v1)** — 565 train / 102 validation / 108 test — remains unchanged. Load **v2** with `load_dataset("gesh75/netops-intent2cli", "v2")`.

Trained model: [gesh75/netops-intent2cli-qwen0.5b-lora](https://huggingface.co/gesh75/netops-intent2cli-qwen0.5b-lora) (experimental). Code: [github.com/gesh75](https://github.com/gesh75) · Docs: [gesh75.github.io](https://gesh75.github.io/)

## v2

- 1215 training rows: 565 inherited v1 + 650 new operator-style synthetic requests.
- Validation 102 and original test 108 are unchanged. Commands are disjoint between training/validation/test, including after whitespace normalization.
- Separate **operator_test** slice: 88 new synthetic paraphrases of original test commands; excluded from all training/checkpoint selection.
- Every new pair has a command copied verbatim from the 70,006-row [multivendor-cli-commands](https://huggingface.co/datasets/gesh75/multivendor-cli-commands) Parquet, with matching vendor and OS. New intents were created offline with assistant-authored task rules/paraphrase grammar; no paid API or copied vendor-document prose.
- New training requests by vendor:

| vendor | n | vendor | n |
|---|---:|---|---:|
| Cisco | 59 | Linux | 26 |
| FRR | 66 | PAN-OS | 23 |
| FortiOS | 57 | Mikrotik | 22 |
| VyOS | 51 | Aruba | 17 |
| Juniper | 50 | Huawei | 8 |
| Wireshark | 47 | NVIDIA | 27 |
| Arista | 44 | Extreme | 37 |
| SONiC | 43 | Nokia | 32 |
| Microsoft | 41 | | |

- `v2/new_pairs.jsonl` is the audit ledger (new training and operator-test requests with explicit split labels), not another training split. **Do not train on operator_test.**
- Source labels: `intent_origin=synthetic-authored-grammar-v2`, `command_origin=verbatim-multivendor-cli-commands`. Inherited v1 rows retain their previous provenance (`corpus` or `curated`).
- New rows pass exact command/vendor/OS membership checks, heuristic near-duplicate rejection, fixed command holdouts and restricted-work-term scanning (zero hits). See the [model card](https://huggingface.co/gesh75/netops-intent2cli-qwen0.5b-lora) for evaluation and training. Adapter-only v2 exact match: 12/108 (11.1%) on the frozen test, 9/88 (10.2%) on `operator_test`. Optional retrieval+validator lookup mode: 63/108 and 58/88 when the gold command is already in the 70,006-row corpus; it cannot invent a command missing from the base. Retrieval-only browser demo: [Intent finder](https://gesh75-multivendor-cli-configurator.static.hf.space/intent-finder.html).

## Limitations

Intents are **synthetic**, not human on-call messages. The operator test shares the author's grammar, so it is not an independent realistic benchmark. Templates/numeric variants remain. The original test contains 16 curated rather than source-corpus-exact commands. Corpus membership is not release-specific correctness: there are placeholders, incomplete grammar entries, unsafe examples and under-specified intents. Never execute these labels without review. The original FortiOS test has only one example; the new operator slice has none.

## Load

```python
from datasets import load_dataset
v1 = load_dataset("gesh75/netops-intent2cli")
v2 = load_dataset("gesh75/netops-intent2cli", "v2")
```

Dataset MIT, unchanged. Model adapter Apache-2.0. Vendor names and command syntax are factual identifiers and trademarks of their respective owners.

---

### GESH · NetOps tools on Hugging Face

| | |
|---|---|
| 📡 **[netlog-ai](https://huggingface.co/spaces/gesh75/netlog-ai)** | Sanitize-first network log analysis — browser lite demo of [netlog-ai](https://pypi.org/project/netlog-ai/) (PyPI 2.1.0) |
| 🛠️ **[multivendor-cli-configurator](https://huggingface.co/spaces/gesh75/multivendor-cli-configurator)** | 70k+ CLI commands across 17 vendors — cheatsheet, N-vendor compare, CLI Studio, Intent finder |
| 🗂️ **[multivendor-cli-commands](https://huggingface.co/datasets/gesh75/multivendor-cli-commands)** | The same 70,006 commands as a Parquet dataset |
| 🧭 **[netops-intent2cli](https://huggingface.co/datasets/gesh75/netops-intent2cli)** | v1 + v2 synthetic operator intents → corpus command pairs |
| 🤖 **[netops-intent2cli-qwen0.5b-lora](https://huggingface.co/gesh75/netops-intent2cli-qwen0.5b-lora)** | LoRA adapter: English intent → vendor CLI (experimental) |

Collection: **[NetOps tools](https://huggingface.co/collections/gesh75/netops-tools-6aef9999054238474f968bcd)** · Code: [github.com/gesh75](https://github.com/gesh75) · Docs hub: [gesh75.github.io](https://gesh75.github.io/)

**Doctrine:** fail closed · sanitize first · evidence leaves, packets don't · honest coverage.