Orvena/orvena-bench

Orvena bench

The Orvena bench measures how well a language model works as the agent of a phone assistant. It has 547 held-out scenarios in 30 languages. Each scenario is a user request together with the data on a simulated iPhone (calendar, contacts, reminders, alarms, places, routes, weather, web pages, files, photos, health and device state) and, when the user has connected services, scripted answers from MCP servers whose tool lists were captured from the real services. A model runs each scenario through the agent loop of the Orvena iPhone app, with the app's system prompt and tool catalog. Every scenario starts at the same local moment, Wednesday 4 November 2026 at 10:30 in the phone's time zone.

A scenario passes only when every check holds:

• the expected tool calls appear in order, each with the expected arguments (other calls in between are allowed, and alternative correct sequences can be listed);
• no forbidden tool is called, or no tool at all when the request needs none;
• the final answer contains the required names, numbers or times and none of the excluded ones;
• the final answer is in the language of the request.

This repository holds the scenarios, the MCP tool lists and tool catalog the bench uses, and the complete recorded runs of three models: Orvena 4B, its base model and Apple's on-device model. It does not include the harness.

Contents

Path	Contents
scenarios/bench.jsonl	The 547 bench scenarios, one JSON object per line
scenarios/restraint.jsonl	60 diagnostic scenarios, 2 per language: 30 requests no tool can do and 30 questions to answer from knowledge, 31 of them with 2 or 3 connections installed. A call to any tool fails them
scenarios/ask.jsonl	30 diagnostic scenarios, 1 per language: requests missing a detail the phone cannot supply. They pass when the model asks instead of acting
fixtures/mcp/<connection>.json	The tools/list result served for each of the 12 MCP connections, 491 tools in all
catalog/orvena-catalog.json	The app's 38 native tools (names, descriptions, JSON schemas, permission metadata), the 12 connections and the base system prompt, exported from the app
results/<set>/<model>.jsonl	Recorded runs, one line per scenario
results/summary.csv, results/by_category.csv, results/by_language.csv	Pass counts computed from the result files

The 547 bench scenarios come from three groups, which the id suffix identifies:

• 407 original scenarios, 13 per language and 30 in English, in 20 categories (62 of them use MCP connections);
• 40 scenarios ending in -h3: places near a named place in another city, multi-step chains, arithmetic without tools, travel times and music requests;
• 100 scenarios ending in -conn: copies of 100 original everyday scenarios with 2 to 6 MCP connections installed that the request does not need. The correct calls do not change.

The categories are calendar, reminders, alarms (alarms and timers), scheduled (tasks the assistant runs later), weather, places, travel, contacts, health, photos, files, web, media, device, chain (two or more dependent tools), recovery (a tool fails or finds nothing), followup (earlier turns in history), no-tool, unsupported and mcp.

The scenarios were written by language-model agents from a brief and the bench's documentation; the 407 original scenarios were written before any training data, and the other sets were added later. No scenario in this repository, including the diagnostic sets, was used to generate or select training data for Orvena 4B.

Results

	Orvena 4B	Base model	Apple on-device
All 547 scenarios	440 (80.4%)	345 (63.1%)	104 (19.0%)
13 languages Apple's model supports (260)	220 (84.6%)	175 (67.3%)	77 (29.6%)
Connections installed, -conn (100)	83	60	23
MCP scenarios (62)	51 (seed 8: 49)	45	8
Restraint, diagnostic (60)	39	41	not run
Asking for a missing detail, diagnostic (30)	8 (seed 8: 11)	18	not run

All counts are recomputed from the files in results/. Runs use seed 7 unless noted; the seed-8 run of Orvena 4B covers only the 62 MCP scenarios and the 30 asking scenarios. Orvena 4B and the base model ran as 4-bit MLX builds (group size 64) with thinking off. The base model is Qwen3.5-4B; see the Orvena 4B model card. Apple's on-device model ran through the app's Apple Intelligence backend in the same loop. It supports 13 of the 30 languages (da, de, en, es, fr, it, ja, ko, nl, pt, sv, tr, vi), and all 547 scenarios were run through it. The runs were made on a Mac with Apple silicon, so seconds is not a phone timing.

results/by_category.csv gives counts per category over the 447 scenarios without the -conn copies, the grouping used in the model card, and results/by_language.csv gives counts per language over all 547.

Reading a scenario

{"id": "en-scheduled-dart-delays", "language": "en", "category": "scheduled",
 "request": "at half 7 tomorrow morning check if there's any delays on the DART and let me know",
 "phone": {"locale": "en_IE", "timeZone": "Europe/Dublin",
           "location": {"place": "Dublin, Ireland", "lat": 53.3331, "lon": -6.2489}},
 "expect": {"calls": [{"tool": "schedule_task", "args": ["{+1d}T07:30", "DART"]}],
            "forbid": ["search_web", "fetch_url", "create_reminder", "schedule_alarm"],
            "answerAny": ["7:30", "07:30"]}}

Field	Meaning
id	<language>-<category>-<slug>, unique across all sets
language	ISO 639-1 code of the request
category	One of the 20 categories above
request	What the user types or says
history	Earlier turns, [{"role": "user" or "assistant", "content": ...}]
connections	Installed MCP connections, matching file names in fixtures/mcp/
phone	The simulated iPhone: locale, timeZone, location, and lists such as events, reminders, alarms, tasks, contacts, places, routes, weather, web, pages, files, photos, health, music, device
tools	Scripted MCP answers, keyed <connection>/<operation>
expect	The checks

Times in phone are relative to the fixed moment ("+1d 10:00" is tomorrow at 10:00, "-2h" two hours ago). Tokens in expectations, MCP rules and phone text are expanded with the run's clock: {+1d} becomes tomorrow's yyyy-MM-dd, {weekday+1d} its weekday name in the phone's language, {iso:+1d 10:00} and {utc:-2h} ISO 8601 timestamps, and {date:+15d} or {datetime:+1d 10:00} a long date in the phone's language. The example above therefore passes when schedule_task is called with tomorrow's date at 07:30 and the word DART in its arguments, none of the four forbidden tools is called, and the answer contains 7:30 or 07:30.

Each entry in tools is a list of rules. The first rule whose match strings all appear in the call's arguments (case- and accent-insensitive), and whose optional nth equals the call's count for that operation, returns its result text, with isError set when error is true and structuredContent from structured. An operation without a matching rule returns a not-found error. Native tools answer from phone in the text formats the app's tools produce, and write operations change the phone for the rest of the turn.

Scoring

Key in expect	Passes when
calls	These calls appear in this order among the calls made
anyOf	At least one of these call sequences appears
args, argsAny, argsGroups, argsNot	Within an expected call: all of args appear in the call's JSON arguments, at least one of argsAny, at least one of each group, none of argsNot
ok	Whether the expected call must have succeeded
forbid	None of these tools was called
noTools	No tool was called
answer, answerAny, answerGroups, answerNot	The final answer contains all of these, at least one, at least one per group, none
language	The answer's language when it should differ from language

All string comparisons are case- and accent-insensitive substring tests. Expected calls count only when they succeeded, except in recovery and unsupported scenarios, and ok overrides this either way. Loader calls (load_skills, load_tools, load_connectors) never need to be listed. An empty answer or a failed turn fails the scenario. The language check removes quoted text and words that occur in tool output, and runs only when at least 12 characters remain. It first settles the languages that Apple's NaturalLanguage recognizer confuses by letters and function words (Ukrainian, Russian and Bulgarian; Slovenian and Croatian; Latvian, Lithuanian, Polish, Estonian and Finnish), and otherwise compares the recognizer's top hypothesis, when its confidence is at least 0.5, with the expected language. Danish and Norwegian, Croatian, Bosnian and Serbian, Indonesian and Malay, and Hindi, Marathi and Nepali count as the same language, and for Estonian, Latvian, Lithuanian and Slovenian requests the recognizer only rejects an answer it reads as English.

Result files

Each line of results/<set>/<model>.jsonl holds one scenario run:

Field	Meaning
id, language, category	The scenario
model, seed	orvena-4b, base-model or apple-on-device, and the run's seed
passed, issues	The verdict and every failed check, as text
answer	The final answer
calls	Every tool call: tool, arguments, failed, and matched (whether a scripted rule or the phone answered it). MCP calls are named <connection>/<operation>
transcript	The turn as the model saw it: user, assistant (with calls) and tool messages, each cut at 4,000 characters. MCP tools appear under the app's internal ids, mcp.<uuid>.<operation>, where the UUID is a hash of the connection id
seconds, reasoningCharacters	Wall time of the turn, and reasoning characters generated (0 with thinking off)
rounds	One entry per model call: the system prompt it received, the declared tool names, the raw model text, the calls, and timings. Empty for Apple's model, whose backend is not recorded at this level

Data preparation and privacy

All people, companies, accounts and documents in the scenarios are fictional; streets, cities and public places are real. After the runs, contact details that could reach a real person or account were replaced, in the same way in the scenarios and in the recorded runs: 145 phone numbers (143 from the scenarios and 2 that models wrote in their answers) now have a national part beginning with 00, 86 email addresses now use the reserved .example top-level domain (for example name@gmail.example), and 4 GitHub user names and 8 GitHub user ids that belong to existing accounts were renamed. Role addresses of large companies, such as airline no-reply senders, were kept. In the 44 addresses that pair a real street with a person (a home, a lease or utility bill, a building's residents, a named doctor or notary, or a dinner at someone's place), the street and city were kept and the house and apartment numbers were replaced with invented ones: house numbers from 9901 upward and apartment numbers from 991 upward. A check over all 1,913 result lines confirmed that the replacement changes none of the substring comparisons the scoring makes, so passed and issues remain valid, but the recorded text differs from what the models saw in these places.

MCP tool lists

Connection	Tools	Source of the tool list
Gmail, Google Drive, Google Calendar	23, 8, 9	Anonymous tools/list requests to Google's hosted servers, October 2026
GitHub	131	The open-source github-mcp-server v2.0.3, run locally, with every toolset and feature-flagged tool
Todoist	47	The open-source @doist/todoist-mcp 13.6.1, run locally
Home Assistant	39	Home Assistant 2026.10.2 with an example home, run locally
Notion, Linear, Airtable, Atlassian, Granola, Zapier	44, 76, 46, 40, 6, 22	Public captures of the hosted servers published by third parties, 2025 to 2026

None of the captures used an Orvena or personal account. The tool names, descriptions and schemas are the services' own; the scripted answers in the scenarios follow the formats these servers use.

Limits

The scenarios are run through the Orvena app's agent loop, which is not published. catalog/ holds the exported native tool definitions and base system prompt, but the code that loads skills, adds runtime tool guidance, routes connections, validates arguments and formats tool results is part of the app. The data is therefore meant for inspection, and for scoring runs made with an equivalent loop; results from a different loop, prompt or tool set are not directly comparable. The checks are rules: a substring check can accept an answer that contains a marker for the wrong reason, or reject a correct answer worded differently, and the language check relies on a statistical recognizer. Each configuration was run once per seed; the two seeds of Orvena 4B on the MCP scenarios (51 and 49 of 62) give an indication of run-to-run variation. Candidate training runs of Orvena 4B were compared on this bench before one was released, so its score here is not an independent test. The diagnostic sets are small.

License

The scenarios, results and summary tables are released under CC BY 4.0. The MCP tool lists in fixtures/mcp/ reproduce the tool descriptions published by each service and remain subject to their owners' terms.

Citation

The bench and the results are described in the technical report of Orvena 4B (PDF).

@misc{orvena_bench_2026,
  title        = {Orvena bench: 547 phone-assistant scenarios in 30 languages},
  author       = {{Orvena}},
  year         = {2026},
  howpublished = {\url{https://huggingface.co/datasets/Orvena/orvena-bench}},
  note         = {Described in the Orvena 4B technical report, \url{https://huggingface.co/Orvena/Orvena-4B}}
}