tool-calls-mini

500 synthetic tool-calling conversations in TRL's conversational format, for supervised fine-tuning. Built to be coherent: every tool result is a plausible function of the arguments it was called with, and every final answer reflects that result — so the set teaches when to call a tool, not just what a call looks like.

Format

Each row has messages and tools. An assistant turn carries tool_calls instead of content; the tool replies as a tool role turn with name and content. The tools column holds JSON schemas generated with transformers.utils.get_json_schema.

{
  "messages": [
    {"role": "user", "content": "How much is 2400 euros in yen?"},
    {"role": "assistant", "tool_calls": [
        {"type": "function", "function": {"name": "convert_currency",
         "arguments": {"amount": 2400.0, "from_currency": "EUR", "to_currency": "JPY"}}}]},
    {"role": "tool", "name": "convert_currency",
     "content": "2400 EUR = 402,168.00 JPY at 167.57 JPY/EUR."},
    {"role": "assistant", "content": "2,400 euros is about ¥402,168 at today's rate."}
  ],
  "tools": [ ... ]
}

Composition

shape	rows	what it teaches
single	232	one call, one result, one answer
sequential	73	a second call whose arguments depend on the first result
parallel	51	several calls in one assistant turn
failure	57	the tool errors and the assistant says so rather than inventing
clarify	44	a required argument is missing, so ask before calling
no_call	43	tools are available and none is the right move

457 rows contain at least one call; 43 deliberately contain none. Those 43 are the point: a dataset of nothing but successful calls teaches a model to reach for a tool at every turn, and one with no error cases teaches it to trust every result.

16 tools across weather, travel, calendar, email, finance, database, translation and logistics. Each row offers the tools it needs plus 1–4 unrelated distractors (mean 3.6 offered), so the right call can't be inferred from the schema list alone. Roughly a quarter of rows carry a system prompt.

Verified

All 500 rows pass TRL's SFT preparation with assistant_only_loss=True — no row is fully masked, and the mask covers the tool_calls turns as well as the final text answer. Mean 698 tokens per row (Qwen3 tokenizer), 12.4% of tokens supervised. No row calls a tool that isn't offered in its own tools column.

Caveats

Synthetic and template-composed: the entity pools are finite, so phrasings recur with different arguments, and the tool results are plausible rather than real. Suited to teaching the tool-calling protocol and call/no-call judgement; not a benchmark, and not a substitute for traces from a real tool-using system.