--- license: apache-2.0 task_categories: - question-answering - text-generation language: - en - zh tags: - agentic-rl - deep-research - sft - cold-start - tool-use - reasoning - rag - trajectories pretty_name: LiteResearcher SFT Cold-Start Trajectories size_categories: - 10K **Distilled deep-research trajectories used for the SFT cold-start of LiteResearcher-4B** [![Paper](https://img.shields.io/badge/Paper-arXiv-b31b1b?logo=arxiv&logoColor=white)](https://arxiv.org/abs/2604.17931) [![Code](https://img.shields.io/badge/Code-GitHub-181717?logo=github&logoColor=white)](https://github.com/simplex-ai-inc/LiteResearcher) [![Webpage](https://img.shields.io/badge/Project-Webpage-0a0a0a?logo=githubpages&logoColor=white)](https://simplex-ai-inc.github.io/LiteResearcher/) [![Model](https://img.shields.io/badge/Model-LiteResearcher--4B-ffcc00?logo=huggingface&logoColor=black)](https://huggingface.co/simplex-ai-inc/LiteResearcher-4B) [![RL Data](https://img.shields.io/badge/RL%20Data-LiteResearcher--Data-ffcc00?logo=huggingface&logoColor=black)](https://huggingface.co/datasets/simplex-ai-inc/LiteResearcher-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`](https://huggingface.co/datasets/simplex-ai-inc/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 ``. > **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`](https://huggingface.co/datasets/simplex-ai-inc/LiteResearcher-Data); > the webpage corpus behind the local search/browse environment lives in > [`simplex-ai-inc/LiteResearcher-Corpus`](https://huggingface.co/datasets/simplex-ai-inc/LiteResearcher-Corpus). > This repo is the *supervised trajectories* that teach the tool-use loop before > any RL happens. --- ## At a glance | | | |---|---| | Rows | 68,231 | | Format | Parquet, 9 shards, ~732 MB total | | Messages per trajectory | min 3, mean 18.3, max 197 | | Total assistant turns | 591,304 | | Language | ~74 % English / ~26 % Chinese (by question) | | Context length | fits a 64 K-token training window (Qwen3 tokenizer) | --- ## Source mix Trajectories are distilled over questions drawn from these upstream pools: | `source` | Rows | What it is | |---|---|---| | `direct_information_seeking_datagen_row1-row42748_v2` | 27,744 | Synthesised direct information-seeking questions (first batch) | | `MiroRL_GenQA` | 10,400 | Generated QA from the MiroRL question pool | | `taskcraft` | 8,645 | TaskCraft-style compositional task questions | | `asearcher` | 8,364 | ASearcher-style search questions | | `direct_information_seeking_datagen_row42748-_v2` | 8,265 | Synthesised direct information-seeking questions (second batch) | | `chinese_QA` | 2,502 | Chinese-language QA | | `BenchSeedQA` | 2,311 | Seed questions derived from benchmark-style tasks | 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 | Column | Type | Description | |---|---|---| | `conversations` | `list[{role, content}]` | The full trajectory in ShareGPT form. Roles are `system`, `user`, `assistant`. | | `source` | string | Upstream question pool (see table above). | **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 `` block with JSON function signatures for `search` and `browse`, and the requirement that the final response be wrapped in `` tags. `assistant` turns emit reasoning followed by either a `` or the final ``. --- ## How to use ### With 🤗 datasets ```python 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 ```bash hf download simplex-ai-inc/LiteResearcher-SFT-Data --repo-type dataset --local-dir ./literesearcher_sft ``` Register it in `data/dataset_info.json`: ```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): ```yaml 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 (`` / ``), 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](https://arxiv.org/abs/2604.17931) 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/`](https://github.com/simplex-ai-inc/LiteResearcher/tree/main/Inference) harness. If you spot a row that shouldn't be public, please open an issue on the [GitHub repo](https://github.com/simplex-ai-inc/LiteResearcher/issues). --- ## License Released under **Apache-2.0**, matching the code and the RL data release. --- ## Citation ```bibtex @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} } ```