kotraj-bench

A Korean agentic tool-use benchmark. 340 tasks across three domains, evaluated against a deterministic mock environment with a fixed user simulator.

Scoring is entirely code — no LLM judge. A task succeeds only if the required tool calls were made, in the required order, with arguments matching the specification, and the environment reached the required terminal state.

Splits

Split	Tasks	Purpose
test	290	Reporting
dev	50	Prompt and harness tuning

Domain	test		Case pattern	test
travel	100		multi_turn	101
calendar	99		sequential	90
commerce	91		parallel	40
			no_call	30
			single	29

no_call tasks are ones where the correct behavior is to not call a tool — out-of-scope requests, or questions answerable directly. See the warning under Results.

Baselines

290 tasks × 3 runs = 870 rollouts per model. All four ran against the same frozen environment and user simulator (env_lock.json).

Model	Params	pass@1	pass^3	pass_any	Avg turns	Tool precision
kotraj-qwen3.5-2B	1.9B	0.297	0.214	0.379	3.22	0.537
Qwen3.5-2B	1.9B	0.241	0.141	0.341	3.74	0.562
LFM2.5-1.2B-Instruct	1.2B	0.223	0.197	0.255	7.01	0.647
gemma-4-E2B-it	~2B	0.190	0.138	0.245	7.39	0.343

pass^3 = correct on all three runs. pass_any = correct on at least one.

McNemar's exact test against Qwen3.5-2B on pass^3: kotraj-qwen3.5-2B p=0.0055, LFM2.5-1.2B p=0.0479, gemma-4-E2B p=1.000.

env_lock.json holds hashes of the schemas, mock implementation, and prompts. Verify it before comparing against these numbers — if any hash differs, you are measuring a different benchmark.

Read no_call separately

Case	kotraj-qwen3.5-2B	Qwen3.5-2B	LFM2.5	gemma-4
single	0.552	0.391	0.471	0.161
multi_turn	0.297	0.195	0.112	0.079
sequential	0.167	0.130	0.044	0.078
parallel	0.217	0.208	0.200	0.133
no_call	0.544	0.633	0.922	1.000

gemma-4 scores 1.000 on no_call and 0.079 on multi_turn. It rarely calls tools at all — its average conversation runs 7.39 turns against kotraj-qwen3.5-2B's 3.22, answering in text instead of acting. A model that never calls a tool scores perfectly on no_call and near zero everywhere else.

Do not report a single aggregate without the case breakdown. no_call rewards inaction, and at 30/290 tasks it moves the headline number.

What a task contains

{
  "id": "s1_20260803-1250_b0069_3",
  "domain": "commerce",
  "case_pattern": "single",
  "task": "기존 주문 ORD0001의 취소 가능 여부를 확인하고 취소 진행을 원한다.",
  "initial_message": "나 예전에 삼다수 시켰던 거 있잖아, ORD0001. 그거 아직 취소 되나?",
  "persona": {"연령대": "30대", "격식": "반말", "스타일": "급함", "지역 맥락": "대전·충청 생활권"},
  "hidden_info": {},
  "success_criteria": {"required_calls": [{"name": "get_order_status", "args_subset": {"order_id": "ORD0001"}}]},
  "tools_subset": ["search_products", "get_order_status", "cancel_order"],
  "seed": 1434659138,
  "split": "test"
}

Field	Description
initial_message	The user simulator's opening turn
persona	Register, urgency, region, digital literacy — drives how the simulator responds
hidden_info	Facts the simulator reveals only when asked
success_criteria	Required calls, argument constraints, ordering, terminal-state conditions
tools_subset	Tools exposed for this task
seed	Mock environment seed — the environment is reproducible from this alone

Contamination control

The benchmark was curated from a task pool disjoint from the training data, then filtered in both directions.

Method	char-5gram MinHash (num_perm=128), Jaccard similarity
Near-duplicate threshold	0.85
Structural + text threshold	0.60
Benchmark tasks checked	340
Training tasks checked	11,744
Collisions	0
Max similarity observed	0.25
Similarity p50 / p90 / p99	0.086 / 0.148 / 0.203

484 candidates were generated; 340 kept. One was dropped for exact text match against training data; the rest were dropped to hold the domain and case-pattern quotas.

MinHash was chosen over embedding similarity deliberately: it is reproducible without a model dependency, so anyone can re-run the audit and get identical numbers. contamination_audit.json contains the full record, including the dropped sample.

A reverse filter also runs during data synthesis, so new training data cannot drift into the benchmark.

Metrics

Metric	Definition
goal_success	All required_calls satisfied, in order, with matching arguments, and terminal state matches
pass@1	Fraction of (task, run) pairs that succeed
pass^k	Fraction of tasks correct on all k runs — reliability, not best-of
tool_precision	Successful calls / total calls
format_violation	Model output could not be parsed as a tool call
sim_fault	Failure attributable to the user simulator, not the agent

sim_fault ran 0.059–0.093 across the four baselines. Treat it as the harness noise floor: differences smaller than that are not meaningful.

Truncation (turn or call cap reached) and repeated format failures count as hard failures regardless of criteria satisfaction. A conversation cut off mid-task is not a success.

Limitations

• Three domains, 30 mock tools. Results do not predict behavior on unseen tool schemas.
• Synthetic tasks. Generated with an Upstage-family model, then filtered by rule gates. Not sampled from real user traffic.
• The user simulator is an LLM. It is pinned to temperature 0 and hashed in env_lock.json, but it is not a real user, and sim_fault is not zero.
• no_call rewards inaction (see above).
• Absolute scores are low. The best baseline reaches pass^3 0.214. This benchmark is not saturated, but it also means small differences sit close to the noise floor.
• Tool-call formats differ by model family. Each model emits calls in the format it was trained on. A harness that assumes one format will score other models as total failures — we measured gemma-4 at 0/50 before fixing the parser. Parse each model's native format.

License

CC-BY-4.0.

Citation

@misc{kotrajbench2026,
  title  = {kotraj-bench: A Korean Agentic Tool-Use Benchmark},
  author = {DONGRYEOL LEE},
  year   = {2026},
  url    = {https://huggingface.co/datasets/drlee1/kotraj-bench}
}

Related

• drlee1/kotraj — the training dataset this benchmark is filtered against
• drlee1/kotraj-qwen3.5-2B — a model fine-tuned on that dataset, evaluated above