Dataset Viewer
Auto-converted to Parquet Duplicate
conversations
listlengths
4
182
system
stringclasses
282 values
tools
stringlengths
205
82.1k
[ { "from": "human", "value": "Hi! I’d like to begin by reviewing the tasks currently assigned to Alice Chen, especially any that are still in the “pending” status. Can you help me identify those tasks?" }, { "from": "gpt", "value": "<think>\nOkay, let's see. The user wants to check tasks assigned...
You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environment by step-by-step use of available tools, while proactively communicating with the user when necessary, until the user ends the conversation. At each step, you will receive either the user's task/reply or the environmen...
[{"type": "function", "function": {"name": "get_task_by_id", "description": "Retrieve full details of a task given its unique task ID.\n\nArgs:\n task_id (str): The unique identifier of the task.\n\nReturns:\n dict: {\n \"success\": True,\n \"data\": TaskInfo # Full task information\n }\n or\...
[ { "from": "human", "value": "Hi! I’d like your help with some task management. Could you help me locate every task that’s currently marked as “pending” and doesn’t have a due date set?" }, { "from": "function_call", "value": "<think>\nOkay, the user is asking to find all tasks that are \"pending...
You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environment by step-by-step use of available tools, while proactively communicating with the user when necessary, until the user ends the conversation. At each step, you will receive either the user's task/reply or the environmen...
[{"type": "function", "function": {"name": "get_task_by_id", "description": "Retrieve full details of a task given its unique task ID.\n\nArgs:\n task_id (str): The unique identifier of the task.\n\nReturns:\n dict: {\n \"success\": True,\n \"data\": TaskInfo # Full task information\n }\n or\...
[{"from":"human","value":"Hi! Let’s start by focusing on the tasks in project P001 (“Website Red(...TRUNCATED)
"You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environm(...TRUNCATED)
"[{\"type\": \"function\", \"function\": {\"name\": \"get_task_by_id\", \"description\": \"Retrieve (...TRUNCATED)
[{"from":"human","value":"Hey! Could you please show me all the tasks whose status isn’t “comple(...TRUNCATED)
"You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environm(...TRUNCATED)
"[{\"type\": \"function\", \"function\": {\"name\": \"get_task_by_id\", \"description\": \"Retrieve (...TRUNCATED)
[{"from":"human","value":"Hi! Could you help me find all tasks that are currently both marked as hig(...TRUNCATED)
"You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environm(...TRUNCATED)
"[{\"type\": \"function\", \"function\": {\"name\": \"get_task_by_id\", \"description\": \"Retrieve (...TRUNCATED)
[{"from":"human","value":"Hey! Could you show me all the tasks from the “Zephyr Migration” proje(...TRUNCATED)
"You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environm(...TRUNCATED)
"[{\"type\": \"function\", \"function\": {\"name\": \"get_task_by_id\", \"description\": \"Retrieve (...TRUNCATED)
[{"from":"human","value":"Hey! I’d like to work on TASK‑003. Could you check its current status (...TRUNCATED)
"You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environm(...TRUNCATED)
"[{\"type\": \"function\", \"function\": {\"name\": \"get_task_by_id\", \"description\": \"Retrieve (...TRUNCATED)
[{"from":"human","value":"Hi! I’d like to start by working with tasks in project P001. Could you h(...TRUNCATED)
"You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environm(...TRUNCATED)
"[{\"type\": \"function\", \"function\": {\"name\": \"get_task_by_id\", \"description\": \"Retrieve (...TRUNCATED)
[{"from":"human","value":"Hey! I need help updating some tasks in our system. Can you help me locate(...TRUNCATED)
"You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environm(...TRUNCATED)
"[{\"type\": \"function\", \"function\": {\"name\": \"get_task_by_id\", \"description\": \"Retrieve (...TRUNCATED)
[{"from":"human","value":"Hey! I’d like help updating some tasks that are currently marked as “p(...TRUNCATED)
"You are a helpful assistant. Your goal is to fulfill the user's requests in an interactive environm(...TRUNCATED)
"[{\"type\": \"function\", \"function\": {\"name\": \"get_task_by_id\", \"description\": \"Retrieve (...TRUNCATED)
End of preview. Expand in Data Studio

Agent World SFT 48K

A supervised fine-tuning dataset containing 44,831 agent-oriented conversations from the agentworld_fixed SFT corpus used for the Qwen3-8B Agent World checkpoint (qwen3-8b-agentworld-fixed-0822).

Dataset structure

The dataset contains one train split. Each example has the following fields:

{
  "conversations": [
    {"from": "human", "value": "..."},
    {"from": "gpt", "value": "..."}
  ],
  "system": "...",
  "tools": "..."
}

conversations uses the ShareGPT format (human, gpt, observation, and function_call roles where applicable). system contains the system prompt, and tools contains the serialized tool definitions.

Decontamination

This release removes 3,776 conversations (7.77% of the original 48,607) whose opening user turn reproduced an opening user turn from the tau2-bench evaluation split.

A conversation was dropped when rapidfuzz.fuzz.ratio between its first human turn and any tau2 evaluation opening was at least 95, after folding whitespace, unifying curly quotes and lowercasing. The 717 reference openings were harvested from tau2 evaluation runs covering all 50 airline, 114 retail and 114 telecom evaluation tasks, which captures the many phrasings the user simulator produces for the same underlying task. All removed rows were from the telecom domain.

Anyone benchmarking on tau2-bench should use this release rather than an earlier revision. The pre-decontamination data remains available in this repository's commit history.

Files

The data is stored as 30 Zstandard-compressed Parquet shards for efficient streaming and compatibility with the Hugging Face Dataset Viewer.

Usage

from datasets import load_dataset

dataset = load_dataset("dongguanting/Agent-World-SFT-48K")
print(dataset["train"][0]["conversations"])

Notes

Users should review the data and apply appropriate filtering and safety checks for their use case.

Downloads last month
13