--- license: cdla-permissive-2.0 language: - en - de - fr - it - sv - ur - hi - ar - gsw task_categories: - text-generation pretty_name: Apertus Lightpath benchmark (operator routing policies for optical networks) --- # Dataset Card: Apertus Lightpath benchmark (operator routing policies for optical networks) Instructions an operator of an optical transport network might type about how a lightpath is routed ("the Zurich–Milan span is under maintenance", "at most four hops, then the shortest in km", "for 400G services, avoid Basel until further notice"), each with the line of a small policy language it means, and the answers of several Apertus readers to each. Built for Hack Apertus (online hackathon, October 2026, Track 2B) by the project *Apertus Lightpath*. ## Dataset Details - **Curated by:** the project's author (Wadood Ranjha), with sentences written by separate AI instances from briefs; see *Source Data*. - **Language(s) (NLP):** English, German, French, Italian; Swedish, Swiss German and Urdu/Hindi in Latin letters in the template lines; the parallel sets add Urdu/Hindi in Latin letters, Urdu in Urdu script, Hindi in Devanagari, and Arabic in Arabic script and in Latin letters (Arabizi). - **License:** CDLA-Permissive-2.0 (the whole dataset and every instance). - **Git repository:** [github.com/theBlackfish01/ApertusNetworkGym](https://github.com/theBlackfish01/ApertusNetworkGym) (`track_2b/`), which holds the generators (`src/corpora3_tests.py`, `src/corpora4_tests.py`, `src/swiss.py`, `src/benchmark_export.py`). ## Uses ### Direct Use Evaluating a model that turns an operator's instruction into a structured routing policy: for each instruction the expected line is given, so a translation can be scored exactly (same meaning, applied with another meaning, not applied). The responses file gives the answers of the project's readers, so new models can be compared with them on the same sentences. ### Out-of-Scope Use The sentences are synthetic: written by AI instances from briefs, or translated by them (the parallel sets), or, for 250 template lines, written outside the project's sentence generators by the author, partly with an AI assistant (ChatGPT 6.1 Sol). They are not a record of what operators say. The policy language carries seven slots; instructions outside it are included only as material the model should refuse or hand on. ### Specifications - `evaluation.jsonl`: one row per instruction: `id`, `set`, `lang`, `kind`, `text`, `topology` (`nobel-eu` or `swiss`), `means` (the line of the fourth format it means, or `unsupported`, `no_instruction`, `unresolved`), `spec` (the same as JSON), `written_by`. - `responses.jsonl`: one row per instruction and reader: `id`, `reader`, `line`, `outcome`. - `metadata.json`: the sets, the readers, the outcome definitions, the line format, counts. ## Dataset Structure Sets (`set` field): the policy language in four languages (`v3_en`, `v3_de`, `v3_fr`, `v3_it`: every clause kind alone, clause pairs seen and never seen in training, triples, two preferences); instruction types shown to no component (`v3_new`, `v3_new_hard`, `v3_new_impossible`); instructions carried in part (`v3_partial`); other routing instructions (`v3_other`); text that is no instruction (`v3_not`); places outside the network (`v3_faulty`); the template lines: the owner set (`owner`, 40) and five sets on the project's template, per request (`template1` … `template5`; 25, 25, 25, 44, 13) and standing policies (`template1_standing` … `template5_standing`; 15, 15, 15, 26, 7); standing policies with a scope or a lifetime (`v4_en` … `v4_it`); the same sentences on a Swiss network with the cities in each language's own name (`v3_swiss_en` … `v3_swiss_it`); and the parallel sets, the same 60 items in nine languages and scripts (`parallel_`, 40 per request, and `parallel__standing`, 20; codes `en`, `de`, `fr`, `it`, `ur_latn` (Urdu/Hindi in Latin letters, written as Roman Urdu), `ur_arab`, `hi`, `ar`, `ar_latn` (Arabic in Latin letters)), with an `item` number shared by the nine versions of an item. The line format: `avoid=; via=; hops=; km=; path=<0-4|any>; prefer=; scope=; until=; rest=`, or `unsupported`, or `no_instruction`. ## Dataset Creation ### Curation Rationale To measure whether a small open model reads operator instructions right, including combinations of clauses it never saw together, languages, names of places, and instructions outside its language, with test sentences written by someone other than the people who wrote the training templates. ### Source Data #### Data Collection The sentences of the `v3_*` sets were written by four separate AI instances (one per language) from a brief that named the clause kinds, the cities and none of the project's wording. The scope and lifetime phrases of the `v4_*` sets were written by another instance from a brief, and attached to `v3` sentences by a script. The `v3_swiss_*` sets fill the `v3` sentences with Swiss cities in each language's own name. The `owner` and `template*` sets were written outside the sentence generators by the author, partly with ChatGPT 6.1 Sol, the `template*` sets on the project's template (`data/instructions/contributor_template.txt`), each line labelled with what was meant and its language. The `parallel_en` items were written by another separate AI instance from a brief (clause kinds, the network, the template's labels), and each other `parallel_*` set by one more instance that translated the English sentences alone, not their labels; no native speaker has checked the translations. Placeholders (cities, links, limits, bit rates, hours) were filled by scripts with values from the topology. #### Source Data Producers and Data Subjects Producers: AI instances (Claude, Anthropic) writing or translating from briefs; the author, partly with an AI assistant (ChatGPT 6.1 Sol), for the 250 template lines. No data subjects: the sentences describe no person, and no data was collected from people. #### Ownership and Consent Management All sentences were created for this dataset by its author or by AI instances under the author's direction; nothing was collected from other sources. #### Data Processing and Quality Control Every expected line is produced by the same script that fills the sentence, so it cannot drift from the sentence; lines are validated against the topology (a place that is not a node makes the instruction `unresolved`); duplicates within a set are dropped. The sets with decisions behind them (`v3_new_hard`, `v3_new_impossible`) were placed on simulated requests by a condition on the instruction and the request only. #### Data Instance Licenses CDLA-Permissive-2.0 for every instance. ### Annotations #### Annotation Process The annotation is the expected line, generated with the sentence (see above). The responses are model outputs, decoded under a grammar that admits only lines of the format; their `outcome` is computed by comparing the meaning of the line with the expected one. #### Annotator Details Scripts in the project repository; the 250 template lines were labelled by the author with the template's words, and the parallel items carry the English item's label (`src/corpora3_owner.py` turns the labels into lines). ## Risks, Bias, and Limitations The sentences are AI-written and may be more regular than what operators type. Four languages only; Swiss German and Romansh are not covered. The policy language carries seven slots; the dataset says nothing about instructions it cannot carry beyond asking that they be handed on or refused. ### Personally Identifiable Information None: the only names are those of cities. ### Sensitive Information None. ### Recommendations Use the `v3_new*` and `v3_partial` sets to measure what a model does with what it was never shown, not only the sets inside the language; report instruction-blind baselines next to any decision-level number. ## Citations [optional] The project's technical report (Hack Apertus online hackathon, October 2026).