tuandunghcmut/deepresearch-9k-tool-calling-strict

DeepResearch-9K — Tool Calling Format (Strict)

Strict converted version of artillerywu/DeepResearch-9K.

Key difference from the standard version:

When an assistant message contains tool_calls, the content field is null. <think> reasoning blocks are dropped from tool-calling turns.

Dataset Summary

Property	Value
Source	artillerywu/DeepResearch-9K
Samples	3,974
Tool	search
Format	OpenAI-compatible messages + tools_json

Difficulty breakdown

Difficulty	Samples
1 (Easy)	826
2 (Medium)	860
3 (Hard)	2,288

Schema

{
    "messages":   str,  # JSON-serialized list of message dicts
    "tools_json": str,  # JSON-serialized list of tool definitions
    "difficulty": int,  # 1 = easy, 2 = medium, 3 = hard
    "source":     str   # "deepresearch-9k"
}

Strict Rule

# Tool call → content is always null, thinking is dropped
{"role": "assistant", "content": null, "tool_calls": [{"function": {"name": "search", ...}}]}

# Final answer → no tool_calls
{"role": "assistant", "content": "Steve Jobs", "tool_calls": []}

Use cases

• SFT for tool-use: Train models to issue search calls directly without verbose reasoning
• Agentic research: Multi-hop web search with parallel query variants
• Difficulty-stratified training: Filter by difficulty field for curriculum learning

Loading

from datasets import load_dataset
import json

ds = load_dataset("tuandunghcmut/deepresearch-9k-tool-calling-strict", split="train")
sample = ds[0]
messages = json.loads(sample["messages"])

# Verify strict rule
for m in messages:
    if m["role"] == "assistant":
        assert not (m.get("content") and m.get("tool_calls")), "Violation!"

# Show tool calls
for m in messages:
    if m.get("tool_calls"):
        for tc in m["tool_calls"]:
            print(tc["function"]["name"], tc["function"]["arguments"][:80])

Source & Citation

Original dataset: artillerywu/DeepResearch-9K

Repository: Applied-Machine-Learning-Lab/DeepResearch-R1