Datasets:

DanielTobi0
/

openresearcher-sft-deep-research-cleaned

qid int64	question string	answer string	num_tool_calls int64	num_turns int64	messages list
0	"What is the grand finalist where the winner is Collingwood among preseason and night series Austral(...TRUNCATED)	\boxed{Hawthorn}	109	331	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)
1	"Identify the method for personalizing image synthesis models to user-provided visual concepts, espe(...TRUNCATED)	"Personalization of image synthesis models to user-provided visual concepts is achieved through opti(...TRUNCATED)	32	100	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)
2	"Four pharaohs of a golden age of ancient Egypt’s pyramid-building dynasty are described as follow(...TRUNCATED)	Sneferu, Khufu, Khafre, Menkaure	22	70	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)
3	"Identify the non-Habsburg imperial figure in Munich with a famed court artist circle, including a F(...TRUNCATED)	Duke Albrecht IV of Bavaria-Munich, Ulrich Füetrer, Erasmus Grasser	80	244	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)
4	"Identify the European legislative body which inspired the naming of the Citizens Party's legislativ(...TRUNCATED)	"The European legislative body that inspired the naming of the Citizens Party's legislative chamber (...TRUNCATED)	76	232	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)
5	Determine if Neil Patrick Harris has narrated the audiobook version of 'Henry Huggins'.	Yes, Neil Patrick Harris narrated the audiobook version of The Henry Huggins Audio Collection.	17	55	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)
6	"Find information about the associate professor at Huaqiao University (华侨大学) who has the ema(...TRUNCATED)	"Zhang Yiwen (张忆文), Associate Professor at the School of Computer Science and Technology, Huaq(...TRUNCATED)	50	154	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)
7	"A doctoral thesis addressing mechanisms behind Earth's self-regulation was published open access in(...TRUNCATED)	\boxed{26,000/6}	112	340	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)
8	"Identify the English reformer whose late-19th-century campaign, with support from a city official, (...TRUNCATED)	Josephine Butler	72	220	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)
9	"Find information about early 1980s Namco arcade maze games that involve underground tunneling. Iden(...TRUNCATED)	"Dig Dug (1982) is the early 1980s Namco arcade maze game involving underground tunneling. The progr(...TRUNCATED)	13	43	[{"role":"system","content":"You are a deep thinking AI that performs research by searching the web.(...TRUNCATED)

End of preview. Expand in Data Studio

OpenResearcher SFT DeepResearch — Parquet Mirror

This is a re-hosted copy of the tool-reasoning SFT deep-research dataset by Aman Priyanshu, itself a cleaned/restructured version of the OpenResearcher Dataset from TIGER-AI-Lab.

Why this repo exists: the source wasn't laid out as ready-to-download Parquet files. This mirror simply stores the data as plain seed_*.parquet files so you can grab the whole dataset or a single segment easily. No changes were made to the content — all cleaning and formatting credit belongs to the original authors below.

Sources

Cleaned/restructured: AmanPriyanshu/tool-reasoning-sft-RESEARCH-openresearcher-dataset-sft-deep-research-agent-data-cleaned
Original: OpenResearcher/OpenResearcher-Dataset

How to Download

Stored as plain Parquet, so you can pull the whole dataset or just a single segment.

This is a private repository — every method below needs an access token. Get one at huggingface.co/settings/tokens, then either run huggingface-cli login once, set export HF_TOKEN=hf_..., or pass token="hf_..." explicitly.

Files

File	Split	Size
`seed_42.parquet` … `seed_57.parquet` (16 files)	`train`	~4.6 GB total
`eval.parquet`	evaluation	~148 MB

Load the full train split (`datasets`)

from datasets import load_dataset

ds = load_dataset(
    "DanielTobi0/openresearcher-sft-deep-research-cleaned",
    split="train",              # all seed_*.parquet
    token="hf_...",             # omit if logged in / HF_TOKEN is set
)
print(ds)

Stream instead of downloading everything

ds = load_dataset(
    "DanielTobi0/openresearcher-sft-deep-research-cleaned",
    split="train",
    streaming=True,
    token="hf_...",
)
for row in ds.take(3):
    print(row["question"])

Download a single segment

from huggingface_hub import hf_hub_download
import pandas as pd

path = hf_hub_download(
    repo_id="DanielTobi0/openresearcher-sft-deep-research-cleaned",
    filename="seed_42.parquet",     # or "eval.parquet"
    repo_type="dataset",
    token="hf_...",
)
df = pd.read_parquet(path)
print(df.shape)

Load one segment directly with datasets:

from datasets import load_dataset

seg = load_dataset(
    "DanielTobi0/openresearcher-sft-deep-research-cleaned",
    data_files="seed_42.parquet",
    split="train",
    token="hf_...",
)

CLI

# a single segment
huggingface-cli download DanielTobi0/openresearcher-sft-deep-research-cleaned \
    seed_42.parquet --repo-type dataset --local-dir ./data

# the whole repo
huggingface-cli download DanielTobi0/openresearcher-sft-deep-research-cleaned \
    --repo-type dataset --local-dir ./data

Schema

Stored as Parquet with native Arrow nested types (messages are list<struct<role: string, content: string>>, not JSON strings).

Column	Type	Description
`qid`	int64	Original question ID
`question`	string	Original question text
`answer`	string	Original ground-truth answer
`num_tool_calls`	int64	Number of tool calls in the conversation
`num_turns`	int64	Total number of turns
`messages`	list[struct]	Multi-turn conversation with roles: system, user, reasoning, tool_call, tool_output, answer

Message format

- system: system prompt with JSON tool schemas
- user: original question
- reasoning: <think>...</think>
- tool_call: <tool_call>{"name": "...", "arguments": {...}}</tool_call>
- tool_output: <tool_response>...</tool_response>
- reasoning: <think>...</think>
- ...
- answer: <answer>...</answer>

Tools: gpt_oss_browser_search, gpt_oss_browser_open, gpt_oss_browser_find.