zake7749/Qwen-3.6-plus-agent-tool-calling-trajectory

---
dataset_info:
features:
- name: id
dtype: string
- name: domain
dtype: string
- name: score
dtype: float64
- name: reward
dtype: float64
- name: num_turns
dtype: int64
- name: messages
list:
- name: content
dtype: string
- name: reasoning_content
dtype: string
- name: role
dtype: string
- name: tool_call_id
dtype: string
- name: tool_calls
list:
- name: function
struct:
- name: arguments
dtype: string
- name: name
dtype: string
- name: id
dtype: string
- name: type
dtype: string
- name: tools
dtype: string
splits:
- name: train
num_examples: 3950
license: apache-2.0
language:
- en
tags:
- tool-calling
- function-calling
- multi-turn
- reasoning
- agent
- sft
- rejection-sampling
size_categories:
- 1K<n<10K
---

# Qwen 3.6 Plus — ToolScale Agent SFT Dataset

This is a subset of multi-turn tool-calling trajectories generated with [Qwen 3.6 Plus](https://openrouter.ai/qwen/qwen3.6-plus) on the [ToolScale](https://huggingface.co/datasets/nvidia/ToolScale).

## Key Stats

| Metric | Value |
|--------|-------|
| Training rows | 3,950 (split by assistant turn) |
| Unique conversations | 582 |
| Total messages | 39,814 |
| Avg turns per conversation | 9.1 |
| Mean score (action match) | 0.568 |
| Mean reward | 0.484 |

## Domains

| Domain | Rows | Description |
|--------|------|-------------|
| bank | 2,227 | Account management, transfers, card freeze/unfreeze |
| ecommerce | 1,119 | Order tracking, returns, payment inquiries |
| basketball | 604 | Game stats, schedules, player information |

## Schema

Each row contains:

| Field | Type | Description |
|-------|------|-------------|
| `id` | `string` | Sample ID |
| `domain` | `string` | Task domain |
| `score` | `float` | Action-matching score against ground truth (0–1) |
| `reward` | `float` | Composite evaluation reward |
| `num_turns` | `int` | Number of assistant turns in the full conversation |
| `messages` | `list[dict]` | Conversation in OpenAI message format |
| `tools` | `string` | JSON string of available tool schemas |

## Example

```python
from datasets import load_dataset

ds = load_dataset("zake7749/qwen-3.6-plus-tool-scale-agent-sft")

sample = ds["train"][0]
messages = sample["messages"] # list of dicts
tools = json.loads(sample["tools"]) # parse JSON string

# Each assistant message has reasoning_content
for msg in messages:
if msg["role"] == "assistant":
print("Reasoning:", msg["reasoning_content"][:200])
print("Content:", msg["content"])
print("Tool calls:", msg.get("tool_calls"))
break
```

A typical conversation looks like:

```
[system] Domain policy + instructions
[user] "Hi, I need to check on my recent order for zip-top bags..."
[assistant] reasoning: "The user is asking for order info. I need to find their account first..."
tool_calls: [find_account_key_by_email(email="[e-mail]")]
[tool] "BERuCRx"
[assistant] reasoning: "Found account key. Now I need to get account details..."
tool_calls: [get_account_details(account_key="BERuCRx")]
[tool] {"account_key": "BERuCRx", "orders": [...], ...}
[assistant] reasoning: "I can see the order details now. Let me summarize for the user..."
content: "Hi Jamie! I found your order. The zip-top bags were charged to..."
```

## Citation

This dataset builds on ToolScale from the ToolOrchestra project:

```bibtex
@article{toolorchestra2025,
title={ToolOrchestra: Elevating Intelligence via Efficient Model and Tool Orchestration},
author={NVIDIA and The University of Hong Kong},
journal={arXiv preprint arXiv:2511.21689},
year={2025}
}
```