tuandunghcmut/toucan-oss

Toucan — OSS split (preprocessed)

Part of the Toucan 1.5M preprocessing pipeline.

Source split	OSS from Agent-Ark/Toucan-1.5M
Rows	457,130
Disk size	~4.3 GB
Source	Open-source model (reasoning-capable) tool-calling conversations

Dataset description

This dataset is the OSS split of Agent-Ark/Toucan-1.5M, preprocessed and normalised to a standard Qwen3 tool-calling format.

All four splits share a unified schema and can be concatenated for training.

Split	Repo	Rows
Qwen3	tuandunghcmut/toucan-qwen3	551,613
SFT	tuandunghcmut/toucan-sft	119,287
OSS	tuandunghcmut/toucan-oss	457,130
Kimi-K2	tuandunghcmut/toucan-kimi-k2	518,516
Total		1,646,546

Schema

Column	Type	Description
messages	string	JSON-encoded list of normalised chat messages
tools_json	string	JSON-encoded list of available tool schemas (OpenAI format), or "" if none
split_name	string	Source split identifier ("OSS")

Message format

Each element of messages follows the Qwen3 / OpenAI tool-calling standard:

// user turn
{"role": "user", "content": "..."}

// assistant text reply
{"role": "assistant", "content": "..."}

// assistant tool call (one or more tools)
{"role": "assistant", "content": "", "tool_calls": [
  {"id": "call_abc123", "type": "function",
    "function": {"name": "tool_name", "arguments": "{...}"}}
]}

// tool result
{"role": "tool", "content": "...", "tool_call_id": "call_abc123"}

Tool schema format

[
  {
    "type": "function",
    "function": {
      "name": "tool_name",
      "description": "What this tool does",
      "parameters": {
        "type": "object",
        "properties": {"arg": {"type": "string", "description": "..."} },
        "required": ["arg"]
      }
    }
  }
]

Source format & normalisation

Original OSS format: Each assistant turn is split into two messages: first carries reasoning_content only (chain-of-thought), second carries the function_call. Uses function_call / function roles. Tools encoded in system message as <tools> XML and in available_tools column.

Normalisation steps applied:

• Extracted tools from available_tools column (preferred) and system <tools> XML.
• Dropped system messages.
• Dropped reasoning_content-only assistant turns (no content, no tool_calls).
• Converted function_call field → standard tool_calls list.
• Converted function role → tool role with tool_call_id.

Drop rate: 0 rows dropped (all 457,130 conversations were valid after normalisation).

Usage

from datasets import load_dataset

ds = load_dataset("tuandunghcmut/toucan-oss", split="train")

# Parse a row
import json
row = ds[0]
messages  = json.loads(row["messages"])
tools     = json.loads(row["tools_json"]) if row["tools_json"] else None

Apply Qwen3 chat template

from transformers import AutoTokenizer

tokenizer = AutoTokenizer.from_pretrained("Qwen/Qwen3-8B")

text = tokenizer.apply_chat_template(
    messages,
    tools=tools,
    tokenize=False,
    add_generation_prompt=False,
)

Load & concatenate all splits

from datasets import load_dataset, concatenate_datasets

splits = ["toucan-qwen3", "toucan-sft", "toucan-oss", "toucan-kimi-k2"]
all_ds = concatenate_datasets([
    load_dataset(f"tuandunghcmut/{s}", split="train") for s in splits
])
print(len(all_ds))   # 1,646,546

Preprocessing pipeline

Source code: preprocess/normalize.py · preprocess/pipeline.py

Agent-Ark/Toucan-1.5M  (raw, split-specific formats)
        │
        ├── Extract tools from available_tools column / system message
        ├── Drop system messages
        ├── Normalize roles  →  user / assistant / tool
        ├── Normalize tool_calls  →  OpenAI format
        ├── Drop verbose assistant turns (reasoning / warm-up)
        └── Trim trailing non-assistant turns
        │
tuandunghcmut/toucan-oss  (unified Qwen3 format)

License

Same as the source dataset — Apache 2.0.