Datasets:

dalek-ai
/

agent-tool-router-eval-fr

_doc string	queries list
Parallel EN/FR queries for qualitative encoder evaluation. Each query lists `expected_concepts`: a returned tool name counts as a hit iff its normalized subtokens contain every token of at least one concept (concept matching is order-free, case-insensitive, separator-insensitive). This is more lenient than exact-name m...	[ { "id": "cancel_order", "en": "cancel my pending order and refund the credit", "fr": "annule ma commande en attente et rembourse le crédit", "expected_concepts": [ "cancel order", "refund order", "cancel pending" ] }, { "id": "slack_message", "en": "send a message t...

agent-tool-router · parallel EN/FR evaluation

50 parallel English/French queries used to evaluate dalek-ai/baseline-v1-desc-hybrid (EN-first) versus dalek-ai/baseline-v1-desc-hybrid-multilingual (50+ languages) on a catalog of 18 671 tools collected from public agent benchmarks (tau-bench, Hermes function-calling-v1, ToolACE).

model	top-3 EN	top-3 FR
`baseline-v1-desc-hybrid` (default, MiniLM-L6)	82%	26%
`baseline-v1-desc-hybrid-multilingual`	82%	54%

Same EN top-3, +28pp on French.

Each query exposes:

id: stable identifier of the underlying concept (e.g. cancel_order).
en: query text in English.
fr: query text in French (manually translated, parallel intent).
expected_concepts: list of concepts. A returned tool name counts as a hit if its normalized subtokens contain every token of at least one concept (order-free, case-insensitive, separator-insensitive). Lenient on naming conventions (send_email / sendEmail / send-email all match the send email concept), strict on semantic content (cancel order will not match modify_pending_order_items).

git clone https://github.com/dalek-ai/agent-tool-router
cd agent-tool-router
pip install -e ".[encoder]"
python -m router.eval.eval_fr_pretrained

The script downloads both pretrained models from the Hub and re-runs the exact same matching logic that produced the table above.

MIT.