ToolUse.1

A tool-calling benchmark that scores correct invocation and correct abstention.

114 cases · 78 requiring a call · 36 requiring abstention or clarification · English-only · 9 tool schemas · 3 calling protocols (JSON, XML, plain-text) · deterministic grading, no LLM judge.

Most tool-use evaluations reward a model for emitting a well-formed call. ToolUse.1 also asks whether it should have called at all: roughly a third of the corpus is prompts where the correct answer is to say something useful and call nothing. A model that calls a tool for every prompt and a model that never calls one score about the same.

Composition

Category	Cases	Category	Cases
calculator	15	negative	30
get_weather	8	clarification	5
translate	8	multi_step	6
convert_currency	8	web_search	8
convert_units	8	set_timer	6
save_note	6	get_time	6

114 English cases. Difficulty is labelled per case (easy / medium / hard) and reported as a breakdown.

Group	What it tests
calculator, get_weather, translate, convert_currency, convert_units, set_timer, save_note, get_time, web_search	Single call, correct tool, semantically correct arguments. Most cases ship all nine tools, so selection is tested against distractors.
negative	No call is correct. Knowledge questions, absent capabilities, explicit no-tool instructions, and a quoted call the user asked to be repeated.
clarification	A required argument is missing. Guessing fails; asking the user passes.
multi_step	History contains prior tool calls and results; the follow-up requires the next call, or a comparison answered without any call.

Schema

{
  "id": "weather-en-001",
  "lang": "en",
  "category": "get_weather",
  "difficulty": "easy",
  "tools": ["calculator", "get_weather", "..."],
  "user": "What's the weather in Berlin?",
  "expect": {"kind": "call", "name": "get_weather", "args": {"city": "Berlin"}}
}

expect.kind is call (with name and canonical args) or no_call (optionally with answer_contains_all / answer_contains_any, case-insensitive substring groups). Multi-step cases add a history array of user / assistant (with tool_name, tool_args) / tool messages.

Metrics

Metric	Definition
Score	passed cases / all cases, with a Wilson 95% interval
Correct call when a tool applies	passing among call cases
Correct abstention	passing among no-call cases
Spurious-call rate	no-call cases where any call was emitted (lower is better)
Tool selected correctly	call cases with the right tool name
Arguments exactly correct	call cases with every expected argument matching
Malformed output rate	call attempts that failed to parse

Arguments are compared semantically, not as strings: expression is evaluated as arithmetic (45+12 equals 45 + 12), numbers compare with a 1e-6 tolerance, and cities, languages, currencies, units and timezones go through alias tables (СПб → saint petersburg, английский → en, $ → USD, Moscow → Europe/Moscow). Tool names are alias tolerant. Extra arguments are tolerated and flagged.

Evaluation harness

The harness is in this repository (tooluse1/), alongside the frozen corpus in data/tooluse1.jsonl and the corpus digest in data/manifest.json.

git clone https://huggingface.co/datasets/HardcoreAI/ToolUse.1
cd ToolUse.1
pip install -r requirements.txt

python -m tooluse1 selftest                       # grader sanity check, no model needed
python smoke.py --model HardcoreAI/Nova-1.1-0.8B --limit 8 --batch-size 8
python -m tooluse1 run --model Qwen/Qwen3.5-0.8B --backend transformers --protocol all --batch-size 8 --out results/
python -m tooluse1 run --model my-model --backend openai --base-url http://localhost:8000/v1 --protocol json

Backends: local transformers checkpoints, any OpenAI-compatible /chat/completions endpoint (vLLM, llama.cpp server, Ollama, OpenRouter), or a scripted file for ablations. Local runs take --batch-size (batch the evaluations into one padded generate call; 8–16 saturates a small GPU) and --dtype (float16 where bfloat16 is unavailable, e.g. a T4 — and report it).

Loading is verified, not assumed: the harness reads back the loading report and refuses an architecture whose weights do not match the checkpoint, because from_pretrained only warns in that case and hands you a randomly initialised model that still writes plausible text.

Reproduction settings: greedy decoding, max_new_tokens=256, the harness system prompt unmodified, no few-shot examples, reasoning mode disabled where the runtime supports it. Use the same protocol set when comparing numbers.

Leaderboard

Results are listed on the ToolUse.1 leaderboard. To submit, run the harness, keep the JSON it writes, and open a discussion here with the model id, protocols, case count, decoding configuration and the raw results file.

The leaderboard seeds itself with harness self-checks only (an oracle runner that reads the expectations, which must score exactly 1.000, and a null runner that never calls a tool). Those rows validate the grader — they are not model results, and they are labelled as such.

Contamination

The corpus is fully public. A model trained on these cases will overfit them. Scores should be read as an upper bound for any model whose authors had access to this repository; models evaluated before publication are the fair comparison. A held-out private split is planned for the 2.x line.

Versioning

Any change to a case bumps the minor version and changes the corpus digest in data/manifest.json, which breaks comparability with earlier scores. Grading changes bump the major version. Always report the digest with a number.

Files

data/tooluse1.jsonl      114 cases, one JSON object per line
data/manifest.json       counts, sha256 of the frozen corpus, harness version
SPEC.md                  full specification: grading rules, limitations, contamination policy
smoke.py                 checkpoint loading check plus a throughput measurement
tooluse1/                harness (cases, tools, protocols, parser, grader, runners, CLI)
tests/                   tests covering parser robustness, grading semantics, corpus invariants

Limitations

See SPEC.md §6. The short version: substring-based answer constraints, wording-sensitive clarification expectations, synthetic tool results in multi-step cases, no private split yet, and one documented ambiguity in the plain-text protocol where a call quoted on its own line is indistinguishable from a real call.

Citation

@misc{hardcoreai_tooluse1,
  title  = {ToolUse.1: a tool-calling benchmark with abstention},
  author = {Hardcore AI},
  year   = {2026},
  url    = {https://huggingface.co/datasets/HardcoreAI/ToolUse.1}
}

License

Apache-2.0. The corpus may be redistributed and used for training; see the contamination note above before comparing numbers.