LiteResearcher — RL Training Data

This dataset contains the two-stage curriculum of question–answer prompts used to train LiteResearcher-4B with on-policy GRPO+TIS, fully against a local search / browse environment. Both stages share the same validation set.

What this is not: the underlying webpage corpus (~32 M records, used by the local search/browse environment) lives in simplex-ai-inc/LiteResearcher-Corpus. This repo is just the task prompts + ground-truth answers that drive the RL rollouts.

---

Splits & sizes

No validation split is bundled. The Wikipedia QA monitoring set we used internally is not a meaningful benchmark on its own. For real evaluation use the standard deep-research benchmarks (GAIA, Xbench-DS, Frames, BrowseComp, HLE, Seal-0, WebwalkerQA) via the Inference/ harness. For a quick sanity loss you can point data.val_files at a 100–500-row slice of stage1/train.parquet.

---

Curriculum design

The two stages implement the difficulty-aware curriculum RL described in §4 of the paper:

Stage 1 — RAG-only warmup (10.4 K prompts)

• All prompts are answerable from a single high-quality web page in the local RAG corpus (the source URL is masked from the agent via extra_info.mask_url so the model has to find it through search + browse tool calls).
• Single homogeneous distribution (data_source = stage2_rag_only) — keeps the signal-to-noise ratio high while the agent learns the tool-use loop.
• Used to bring a SFT cold-start (Qwen3-4B + 68.2 K distilled deep-research trajectories) up to a 32 K-context plateau (~step 220, 64.7 % avg).

Stage 2 — Difficulty + diversity ramp (16.2 K prompts)

• 25 buckets across 6 question families (rag_direct / rag_direct_wiki / rag_mqa / rag_mqa_subgraph{5,6,7} / rag_science) crossed with 4 length buckets (lt16k / 16k-32k / 32k-48k / 48k-64k).
• Each bucket is pre-filtered by pass@8 from the Stage-1 checkpoint to keep only problems that are learnably hard (excludes trivial pass and impossible).
• Mixes single-hop, multi-hop (subgraph), and scientific QA so the curriculum spans both reasoning structure and context length.

---

Schema

Each row is a verl-compatible prompt with these fields:

The validation split has one extra url field at the top of extra_info — it is also a gold reference URL, kept for backward compatibility with our eval harness.

---

How to use

With 🤗 datasets

from datasets import load_dataset

# Stage-1 warmup
stage1 = load_dataset("simplex-ai-inc/LiteResearcher-Data", "stage1")
print(stage1)              # DatasetDict: train (10398)
print(stage1["train"][0])  # one example

# Stage-2 curriculum (default config)
stage2 = load_dataset("simplex-ai-inc/LiteResearcher-Data", "stage2")

With the LiteResearcher training pipeline

Both stages are drop-in for the launchers shipped in simplex-ai-inc/LiteResearcher under Training/:

# Download once
hf download simplex-ai-inc/LiteResearcher-Data --repo-type dataset --local-dir ./literesearcher_data

# Stage 1 (single node, 8 × H20)
export TRAIN_DATA=./literesearcher_data/stage1/train.parquet
# Use a slice of train for sanity validation, or supply your own eval parquet:
export VAL_DATA=./literesearcher_data/stage1/train.parquet
bash Training/examples/sglang_multiturn/search_browser/stage1_rag_only.sh

# Stage 2 (2 nodes × 8 H20, resumed from Stage-1 step ~220)
export TRAIN_DATA=./literesearcher_data/stage2/train.parquet
export VAL_DATA=./literesearcher_data/stage2/train.parquet
bash Training/examples/sglang_multiturn/search_browser/stage_2_mix_rag_on_policy_48k.sh

---

How it was built

1. Seed QA generation — A small expert LLM was prompted over the LiteResearcher-Corpus webpages to synthesise question–answer pairs for each page (single-hop) and for randomly sampled subgraphs of the citation graph (multi-hop, subgraph sizes 5/6/7).
2. LLM-judge filtering — Each candidate QA was screened by an LLM judge for answer uniqueness, evidence sufficiency, and freedom from time-sensitive ambiguity.
3. Difficulty bucketing (Stage 2 only) — Surviving prompts were run through the Stage-1 checkpoint at pass@8; only items with a non-trivial, non-saturated pass-rate enter the Stage-2 mix. Items are also bucketed by the realised rollout context length (lt16k / 16k-32k / 32k-48k / 48k-64k) so the curriculum can up-weight harder lengths over time.
4. URL masking — For every row, the originating webpage URL is recorded in extra_info.mask_url; the rollout environment removes this URL from search results so the model cannot one-shot the answer by re-finding the source.

See §3 ("Co-construct Training Data & Corpus") and §4 ("Difficulty-Aware Curriculum RL") of the paper for details.

---

Languages

The dataset is mostly English with a meaningful Chinese subset (especially in the multi-hop / institutional-knowledge buckets of Stage 2). All ground-truth answers are short factual strings; questions can be either language.

---

License

Released under Apache-2.0, matching the code release.

---

Limitations & responsible use

• Public-info QA pairs may include public-facing email addresses, phone numbers, and institutional URLs that exist in the source webpages (e.g., "What is the official contact email of [public institution X]?"). These were already indexed on the open web before scraping. If you spot a row that shouldn't be public, please open an issue on the GitHub repo and we'll patch it.
• The dataset is not balanced for any specific evaluation domain beyond the curriculum design described above. Downstream evaluation should be done on held-out benchmarks (GAIA, Xbench-DS, Frames, BrowseComp, etc.), not on the validation split here (which is single-hop Wikipedia and is only used to monitor training progress).
• All examples assume a working local RAG + browse environment — see simplex-ai-inc/LiteResearcher-Corpus and Environment/ in the code repo to reproduce.

---

Citation

@article{li2026literesearcher,
  title   = {LiteResearcher: A Scalable Agentic RL Training Framework for Deep Research Agent},
  author  = {Li, Wanli and Qu, Bince and Pan, Bo and Zhang, Jianyu and Liu, Zheng and Zhang, Pan and Chen, Wei and Zhang, Bo},
  journal = {arXiv preprint arXiv:2604.17931},
  year    = {2026}
}