aliarda/jev_turkish_tool_call_traces

JEV Turkish Tool Call Traces

Traces of Jev (jev-1.13.0, TypeSafe System One) answering Turkish tool-selection questions from a Turkish tool-calling benchmark.

For each user turn the conversation transcript is the state, and the model answers one choice question: which tool should the assistant call next (or no_tool). The shared set of candidate tools lives in meta/criteria.json.

Columns (train)

column	description
index	row order (0-based)
id / turn_index	conversation id / message index in the source
category	single, parallel, or no_tool
expected	gold tool name(s)
correct	whether jev's choice is in expected
jev_choice / jev_confidence / jev_top5 / jev_probabilities_json	jev's answer
model	model version that answered
input_tokens / output_tokens / elapsed_ms / request_id	usage
user_message / state	the evaluated question and its context

Results

• Overall: 246/308 = 79.87%
• single tool: 199/213
• parallel tools: 10/12
• no tool: 37/83

Notes

• parallel turns accept any of the gold tools as correct (the question is single-choice).
• no_tool is the correct answer when no tool call is expected.
• meta/ holds the gold labels, the candidate-tool set, and the metrics.
• Built from conversational-turkish-tool-call.json.