orbit-ai/orbit-stage-2-27k

⚠️ This is an unverified ORBIT dataset, only the Stage-2, and may contain data inaccuracies. The verified ORBIT dataset to use is orbit-ai/orbit-20k.

ORBIT: A Synthetic Training Dataset for Search Agents

ORBIT is a reasoning-intensive synthetic dataset with complex queries used for training search agents, generated without relying on any paid API services or manual annotation.

---

Overview

Training data for deep search — tasks requiring multi-step retrieval and reasoning over the web — is scarce. Existing datasets either target simple single-hop questions (NQ, HotpotQA) or require costly prerequisites such as knowledge-graph construction or human annotators.

ORBIT addresses this with a fully automatic, four-stage pipeline that generates reasoning-intensive question–answer pairs with short, verifiable answers grounded in real web evidence. The entire dataset was constructed using a frugal setup (a single laptop + free chat interfaces) at near-zero cost.

Key properties:

• 20,000+ verified question–answer training pairs
• 15 diverse domains (TV & Movies, Science & Technology, Art, History, Sports, Music, Video Games, Geography, Politics, Medicine, Finance, Law, Mathematics, Puzzles, Code)
• 4–6 reasoning steps required per question on average — harder than InfoSeek (3–5) and HotpotQA (2)
• Short, verifiable answers averaging 3.5 tokens
• 4.36 verification URLs per pair, spanning Wikipedia, NIH, ScienceDirect, and many more sources
• No paid APIs, no knowledge graphs, no human annotators required

---

Dataset Statistics

Statistic	Value
# Question–Answer pairs	20,147
Avg. question length (tokens)	63.87
Avg. answer length (tokens)	3.46
Avg. reasoning steps	4.42
Avg. verification URLs per pair	4.36
— Non-Wikipedia URLs	3.35
— Wikipedia URLs	1.02
Domains	15

---

Dataset Schema

Each row in orbit-ai/orbit-20k contains the following fields:

Field	Type	Description
_id	string	Unique 32-character hex identifier
question	string	Reasoning-intensive, multi-hop question (86–701 chars)
answer	string	Short, verifiable ground-truth answer (1–191 chars)
verification_checklist	list[dict]	Atomic sub-claims with citation indices required to verify the answer (DeepSeek V3.1)
citation_urls	dict	Mapping of citation index → URL used for verification (DeepSeek V3.1)
category	string	Wikipedia category of the seed page
topic	string	One of 15 high-level domains
scraped_documents	list[dict]	Text scraped from the verification URLs (Scraped using Trafilatura)
predicted_answer	string	Answer produced by the external LLM judge during Stage 4 (Qwen/Qwen3-4B-Instruct-2507 or gpt-oss-120b)
judge_output	string	Stage 4 external judge reasoning and verdict (Qwen/Qwen3-4B-Instruct-2507 or gpt-oss-120b)
external_judge	string	Model used as the external judge (Qwen/Qwen3-4B-Instruct-2507 or gpt-oss-120b)
deepseek_judge	string	Full Stage 3 self-verification trace (DeepSeek-V3.1)
seed	string	Wikipedia page title used as the generation seed

---

Construction Pipeline

ORBIT is built with a four-stage, fully automatic pipeline:

Stage 1: Seed Creation
  └── 15 domains → 100 Wikipedia categories each (via OpenAI Deep Research)
      └── MediaWiki API → unique Wikipedia page titles as seeds

Stage 2: Multi-Hop Q&A Generation
  └── DeepSeek-V3.1 (chat.deepseek.com, DeepThink + Search enabled)
      └── Generates inverted question + short answer from each seed
          → 44,100 raw pairs

Stage 3: Self-Verification
  └── DeepSeek-V3.1 re-verifies each pair with independent web search
      └── Qwen3-4B-Instruct filters pairs with insufficient evidence
          → 27,100 pairs (61.5% pass rate)

Stage 4: External Verification
  └── Round 1: Qwen3-4B-Instruct judges against scraped URL content
  └── Round 2: gpt-oss-120b re-judges all failures from Round 1
  └── Manual review of borderline cases
          → 20,147 verified pairs

Questions are designed to be inverted (easier to verify given the answer) following the BrowseComp methodology, and explicitly require cross-source, multi-hop evidence.

---

Ethical Considerations

ORBIT is built from publicly accessible sources (Wikipedia categories and publicly retrievable webpages). We do not intentionally collect private or personally identifying information. The dataset is intended for factual question answering research only.

Because the pipeline relies on live web content and LLM-based generation, the data may inherit biases or framing effects from source webpages and generation models. Some answers may also become temporally brittle as web content evolves. We recommend treating ORBIT as a versioned research snapshot rather than an authoritative source of truth.

Browser automation (Selenium) was used to interact with DeepSeek Chat under a single authenticated personal account for academic, non-commercial purposes, with sequential rate-limited requests equivalent to manual usage.

---

Citation

If you use ORBIT in your work, please cite our preprint on arXiv:

@misc{thakur2026orbit,
      title={ORBIT: Scalable and Verifiable Data Generation for Search Agents on a Tight Budget}, 
      author={Nandan Thakur and Zijian Chen and Xueguang Ma and Jimmy Lin},
      year={2026},
      eprint={2604.01195},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2604.01195}, 
}

---

Links

Resource	URL
Paper	https://arxiv.org/abs/2604.01195
Dataset	https://huggingface.co/datasets/orbit-ai/orbit-20k
Model	https://huggingface.co/orbit-ai/orbit-4b-v0.1
Hugging Face	https://huggingface.co/orbit-ai
GitHub	https://github.com/castorini/orbit