LiteResearcher — SFT Cold-Start Data

This dataset contains the 68,231 multi-turn deep-research trajectories used to train the SFT cold-start checkpoint that RL (GRPO+TIS) is later launched from — the "68.2 K distilled deep-research trajectories" referenced in the paper and in LiteResearcher-Data.

Each row is a complete ReAct-style episode: a research question, the model's interleaved thinking and search / browse tool calls, the observations that came back, and a final <answer>.

Where this sits in the pipeline: SFT cold-start → Stage-1 RAG warmup → Stage-2 curriculum RL. The RL prompts live in simplex-ai-inc/LiteResearcher-Data; the webpage corpus behind the local search/browse environment lives in simplex-ai-inc/LiteResearcher-Corpus. This repo is the supervised trajectories that teach the tool-use loop before any RL happens.

---

At a glance

---

Source mix

Trajectories are distilled over questions drawn from these upstream pools:

The two direct_information_seeking_*_v2 shards are the same generator run split by row range — treat them as one 36,009-row pool if you want a coarser grouping.

---

Schema

On the role encoding: tool observations (search results, fetched page content) are carried in user turns rather than a separate observation / tool role — this is the ShareGPT convention the LLaMA-Factory recipe expects. A trajectory alternates assistant (reasoning + tool call) ↔ user (observation), ending on an assistant turn that emits the final answer.

The system turn carries the deep-research system prompt: a role description, the <tools> block with JSON function signatures for search and browse, and the requirement that the final response be wrapped in <answer></answer> tags. assistant turns emit reasoning followed by either a <tool_call> or the final <answer>.

---

How to use

With 🤗 datasets

from datasets import load_dataset

ds = load_dataset("simplex-ai-inc/LiteResearcher-SFT-Data", split="train")
print(ds)                          # 68,231 rows
print(ds[0]["source"])
print(ds[0]["conversations"][0])   # system prompt (tools block)
print(ds[0]["conversations"][1])   # the research question

With LLaMA-Factory

hf download simplex-ai-inc/LiteResearcher-SFT-Data --repo-type dataset --local-dir ./literesearcher_sft

Register it in data/dataset_info.json:

{
  "literesearcher_sft": {
    "file_name": "literesearcher_sft/data",
    "formatting": "sharegpt",
    "columns": { "messages": "conversations" },
    "tags": {
      "role_tag": "role", "content_tag": "content",
      "user_tag": "user", "assistant_tag": "assistant",
      "observation_tag": "observation", "system_tag": "system"
    }
  }
}

Cold-start config we used (Qwen3-4B-Thinking-2507, 8×H100, full fine-tune):

dataset: literesearcher_sft
template: qwen3
enable_thinking: true
cutoff_len: 65536          # 64K training window
learning_rate: 2.0e-5
num_train_epochs: 1.0
per_device_train_batch_size: 2
gradient_accumulation_steps: 8
lr_scheduler_type: cosine
warmup_ratio: 0.1
weight_decay: 0.01
bf16: true
gradient_checkpointing: true
flash_attn: fa2
enable_liger_kernel: true
deepspeed: examples/deepspeed/ds_z2_config.json

---

How it was built

1. Question pooling — questions were collected from the sources above, spanning direct information-seeking, multi-hop, compositional (TaskCraft), and Chinese-language QA.
2. Trajectory distillation — a stronger teacher model rolled out full ReAct episodes against the search / browse environment, producing interleaved reasoning, tool calls, and observations.
3. Cleaning — trajectories were filtered for a well-formed output contract (<tool_call> / <answer>), consistent tool-call syntax, answer agreement with the reference, and removal of degenerate repeated-action episodes.
4. Length check — remaining episodes were tokenised with the Qwen3 tokenizer and verified against the 64 K context window.

See §3 and §5 of the paper for how the cold-start checkpoint feeds into the two RL stages.

---

Limitations & responsible use

• Trajectories are teacher-distilled, not human-verified. Intermediate reasoning can contain factual errors, dead-end searches, or hallucinated intermediate claims even when the final answer matches the reference. Treat this as behaviour-cloning data for the tool-use loop, not as a source of ground truth.
• Observations are snapshots of live web content from the collection period (late 2025). Some pages will have changed or disappeared; some contain public-facing contact details that were already indexed on the open web.
• No held-out split is bundled. Evaluate on standard deep-research benchmarks (GAIA, Xbench-DS, Frames, BrowseComp, HLE, Seal-0, WebWalkerQA) via the Inference/ harness.

If you spot a row that shouldn't be public, please open an issue on the GitHub repo.

---

License

Released under Apache-2.0, matching the code and the RL data release.

---

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}
}