Theo

By Openwist. Theo is an open engineering assistant. These datasets teach small models to do engineering work by calling deterministic tools instead of guessing, to ask when a value is missing, to report assumptions and warnings, and to refuse to defeat safety protections.

36,349 records in three subsets

subset	topic	records	train / validation / test	types
ev	Electric vehicles	15,000	13,000 / 1,000 / 1,000	calc 2,382, clarify 1,280, code 723, concept 719, critique 1,502, doc 614, flag 1,485, refuse 912, sim 2,777, toolchain 2,606
humanoid	Humanoid robots	6,349	5,110 / 612 / 627	calc 2,200, clarify 547, concept 96, flag 1,200, refuse 532, toolchain 1,774
code	Engineering code	15,000	13,000 / 1,000 / 1,000	clarify 594, code 14,000, refuse 406

from datasets import load_dataset
ev = load_dataset("Openwist/Theo", "ev")
humanoid = load_dataset("Openwist/Theo", "humanoid")
code = load_dataset("Openwist/Theo", "code")

Format

See DATASET_FORMAT.md. Use system_prompt.txt as the system message and pass the tool schemas in schemas/ through the chat template's tool slot. Tool-call arguments and tool content are JSON strings:

import json
rec = ev["train"][0]
for m in rec["messages"]:
    for call in m.get("tool_calls") or []:
        args = json.loads(call["arguments"])
    if m["role"] == "tool":
        result = json.loads(m["content"])   # includes warnings and assumptions

Tools

EV: unit_check, ev_pack_size, ev_range, ev_motor_torque, ev_gear_ratio, ev_charge_time, battery_ecm_sim, vehicle_longitudinal_sim, thermal_node_sim. Humanoid: hum_joint_torque, hum_actuator_select, hum_com_zmp, hum_battery_runtime, plus unit_check. Code records were executed in a sandbox.

How it was built

Scenarios are sampled with realistic inputs, the real tools are run to produce calls and results, and answers are composed from hand-written fragments. Every record is then verified by code (see DATASET_FORMAT.md), and no validation or test record repeats a train user message once numbers are masked. The same task families (for example refusals and clarifying questions) appear in every split, so scores measure new values and wording variants within known task families, not new task families.

The humanoid subset came from a template generator. The EV and code subsets were supplied as finished data and were cleaned and re-checked (see CLEANING_REPORT.md). The generator code is not published with this dataset.

Limits

• Wording is template-based, so it is narrower than natural user text. Mix in other data for conversational variety.
• Outputs are first-pass estimates for early design, not design sign-off. Numbers come from the tools, not from experiments.
• EV critique and doc records are short: a target check and a brief structured report.
• The humanoid subset has no MuJoCo simulation records and no code, document or critique records.
• Code refactor records are checked by matching printed output, not by an equivalence test.

Intended and unsupported uses

Intended: training and evaluating small language models to route engineering questions to deterministic tools, ask for missing values, and report assumptions and warnings; research on tool-grounded answers.

Unsupported: design sign-off, certification or regulatory evidence; controlling or tuning a real vehicle, battery or robot; safety-critical or medical decisions; any claim that the numbers are validated on real hardware or on Kenyan vehicles or roads. Openwist asks that the data not be used to build weapons or to defeat safety protections. This is a request, not a licence condition: the licence is CC BY 4.0.

Evidence state

Numbers in the records were computed by deterministic software libraries, which makes them Simulated in the sense of the Openwist Safety Framework. The libraries have unit tests, and the tool-using records were checked by code (tool replay, number match, units, warnings). The data has not been Validated against real vehicle or robot measurements, and it contains no Kenyan field data. See RELEASE_RECORD.md for the checks, their dates and the open items.

Safety

Includes refusal examples for requests to defeat safety protections (battery protection, interlocks, thermal cutoffs, current limits, emergency stops). They give no procedure and offer a safe alternative. A model trained on this data is not a substitute for qualified engineering review.

Licence

Creative Commons Attribution 4.0 International (CC BY 4.0). You may use, share and adapt this dataset for any purpose, including commercial use, as long as you give credit to Openwist. See ATTRIBUTION.md.

Citation

Openwist. Theo. 2026. https://openwist.com