Datasetpublic

junlin-modal/k3-toolcall-turn1

k3-toolcall-turn1 191,072 deduplicated tool-calling contexts, normalized to the OpenAI messages + tools schema. Each row is a conversation prefix that stops right before the first assistant turn — a context ready for a model to generate at, not a finished conversation. Built to generate on-policy targets for speculative-decoding draft training: a target model generates at each context, and the draft learns to copy that model's next tokens. The dataset ships contexts, not labels.… See the full description on the dataset page: https://huggingface.co/datasets/junlin-modal/k3-toolcall-turn1.

sourceHugging Facecc-by-4.0updated 2mo agoView on Hugging Face

0likes23downloads

Dataset Card

k3-toolcall-turn1

191,072 deduplicated tool-calling contexts, normalized to the OpenAI messages + tools schema. Each row is a conversation prefix that stops right before the first assistant turn — a context ready for a model to generate at, not a finished conversation.

Built to generate on-policy targets for speculative-decoding draft training: a target model generates at each context, and the draft learns to copy that model's next tokens. The dataset ships contexts, not labels.

Composition

source	rows	upstream license	what it adds
Agent-Ark/Toucan-1.5M (Kimi-K2)	80,000	Apache-2.0	real MCP-server tools (2,000+), agentic queries
minpeter/xlam-function-calling-60k-parsed	59,616	CC-BY-4.0	single-turn calls, many parallel-call examples
glaiveai/glaive-function-calling-v2	39,664	Apache-2.0	volume, "tools present but answer directly" cases
Team-ACE/ToolACE	10,692	Apache-2.0	high API diversity (16k+ tools)
NousResearch/hermes-function-calling-v1	1,100	Apache-2.0	multi-turn reasoning + tool use
total	191,072		24,818 distinct tools

License is CC-BY-4.0 (the most restrictive among the sources; xlam is CC-BY-4.0, the rest Apache-2.0). Commercial use is allowed with attribution.

Format

One JSON object per line:

json

{
  "id": "toucan/12#t1",
  "source": "toucan",
  "tools": [ { "type": "function", "function": { "name": "...", "parameters": { ... } } } ],
  "messages": [ { "role": "user", "content": "..." } ],
  "n_tools": 6,
  "has_system": false,
  "reference_output": { "content": "", "tool_calls": [ { "name": "...", "arguments": "{...}" } ] }
}

—messages — the context. Has at least one user message and stops before the first assistant turn.
—tools — OpenAI-format declarations. Every tool called later in the source conversation is declared here.
—reference_output — the source dataset's original first assistant turn, kept for reference only. Another model wrote it; it is not a training label.

Construction & validation

Source conversations were parsed out of four different in-band formats (glaive's ASSISTANT: <functioncall> text, Hermes' <tool_call> XML, ToolACE's [Fn(arg=...)] strings, Toucan's legacy function_call field with a Kimi-K2 template system message) into a single messages + tools schema. A conversation was dropped, never repaired, when a tool result could not be matched to a preceding call, a call named an undeclared tool, or a payload failed to parse. Sampled rows all render under apply_chat_template(messages, tools=tools, add_generation_prompt=True) on a production tokenizer.

Not included: Salesforce/APIGen-MT-5k (CC-BY-NC-4.0, non-commercial) and Salesforce/xlam-function-calling-60k (gated; the ungated minpeter re-parse is used instead).

Attribution

Derived from the datasets listed above; please cite their original authors (Agent-Ark, Salesforce/minpeter, GlaiveAI, Team-ACE, NousResearch).