jddqd/toolace-sft

ToolACE, prepared for SFT

Derived from Team-ACE/ToolACE (revision 6bda777c88d21e5a204703c1ee45597a8fa4f734), licensed Apache-2.0 by Team-ACE. Please cite the original ToolACE work when using this data.

Each line is a conversation, single- or multi-turn (tool responses, text answers when no function applies or a parameter is missing): messages in the Hugging Face chat format, tools in the exact function-doc format the BFCL evaluation puts in the prompt. Usable with tokenizer.apply_chat_template(messages, tools=tools):

{
  "messages": [
    {
      "role": "user",
      "content": "What is the weather in Paris?"
    },
    {
      "role": "assistant",
      "content": "",
      "tool_calls": [
        {
          "type": "function",
          "function": {
            "name": "get_weather",
            "arguments": {
              "city": "Paris"
            }
          }
        }
      ]
    }
  ],
  "tools": [
    {
      "name": "get_weather",
      "description": "Get the current weather of a city. Note that the provided function is in Python 3 syntax.",
      "parameters": {
        "type": "dict",
        "properties": {
          "city": {
            "type": "string",
            "description": "The city name."
          },
          "unit": {
            "type": "string",
            "description": "celsius or fahrenheit.",
            "default": "celsius"
          }
        },
        "required": [
          "city"
        ]
      }
    }
  ]
}

Processing

1. Tools parsed from the system prompt and normalized to BFCL's function-doc format (the system prompt itself is dropped: the chat template renders the tools as BFCL does): BFCL's Python-3 hint appended to descriptions, BFCL type names (int → integer, nullable unions such as ["integer", "null"] → integer). Rows using another tool format (ToolACE's format augmentation: other key names, custom call syntaxes), non-nullable unions, or two tools with the same name are dropped.
2. Python-like calls ([Tool Name(arg="value")]) parsed into structured tool_calls and validated against the tools: no unknown function, no unknown argument, no missing required argument. Conversations must start with a user message and end with an assistant answer.
3. Unicode NFC normalization and whitespace stripping.
4. Exact deduplication on (question, answer). Decontaminated against the BFCL questions: dropped rows whose user message is a BFCL question or shares 8-grams covering at least 50% of the shorter text.
5. Tool order shuffled (in the source, the called tool is listed first in 48% of rows with several tools, 29% expected by chance).
6. Rows longer than 2048 tokens with the Qwen/Qwen3-0.6B-Base chat template dropped (never truncated).
7. Train / validation split (3%) stratified by BFCL-style category, seed 0.

step	rows
source rows	11,300
valid (converted + calls checked)	10,761
after exact dedup (question + answer)	10,758
after decontamination against BFCL	10,753
within max length	10,666
train	10,346
validation	320

rejection reason	rows
nonstandard_format	518
too_long	87
union_type	20
contaminated	5
duplicate	3
invalid_schema	1

rejected.jsonl contains every dropped source row with its reason.

Statistics

category	train	validation
multiple	2,842	88
parallel_multiple	2,000	62
parallel	1,923	59
no_call	1,491	46
simple	1,395	43
multi_turn	695	22

Train tokens (Qwen/Qwen3-0.6B-Base chat template): mean 682.3, median 624.0, p99 1741.53, max 2047, total 7,058,903.