Datasetpublic

guinansu/literesearcher-stage2-masked

LiteResearcher Stage-2 (URL-masked, train/val split) Stage-2 RL data for training a multi-turn search agent, prepared for verl-style GRPO training with search and browse tools. Provenance Derived from simplex-ai-inc/LiteResearcher-Data (Apache-2.0). Upstream's pipeline is SFT cold-start → Stage-1 RL → Stage-2 RL; this repository carries only the Stage-2 portion. Processing applied on top of upstream: Stage-2 rows only (upstream Stage 1 is a pure local-RAG warmup… See the full description on the dataset page: https://huggingface.co/datasets/guinansu/literesearcher-stage2-masked.

sourceHugging Faceapache-2.0updated 1mo agoView on Hugging Face

0likes15downloads

Dataset Card

LiteResearcher Stage-2 (URL-masked, train/val split)

Stage-2 RL data for training a multi-turn search agent, prepared for verl-style GRPO training with search and browse tools.

Provenance

Derived from `simplex-ai-inc/LiteResearcher-Data` (Apache-2.0). Upstream's pipeline is SFT cold-start → Stage-1 RL → Stage-2 RL; this repository carries only the Stage-2 portion.

Processing applied on top of upstream:

1.Stage-2 rows only (upstream Stage 1 is a pure local-RAG warmup and is not included)
2.URL masking in tool responses, via upstream's merge_rag_with_mask_url.py (extra_info.mask_url)
3.A held-out validation split, stratified to match the training distribution

Upstream Stage 2 ships 16,199 rows; the row-count difference here comes from the masking/split step.

Size

split	rows	size
train	15,687	11.3 MB
validation	512	0.3 MB

Structure

field	description
`question`	the user query
`data_source`	curriculum bucket, `literesearcher_rag_{task}_1-7_{ctx}` — also the grouping key for per-bucket eval metrics
`prompt`	chat-format prompt handed to the policy
`ability`	always `search`
`reward_model`	`{style: "llm", ground_truth: {target: [...]}}` — scored by an LLM judge
`extra_info`	`index`, `split`, `mask_url`, `need_tools_kwargs`, `tools_kwargs` (per-sample `search`/`browse` config)
`metadata`	unused (null)

Curriculum buckets

25 buckets = 7 task types x 4 context-length bands. data_source is what per-subset accuracy is reported against during validation, so each bucket gets its own score rather than a single pooled number.

Task types: direct, direct_wiki, mqa, mqa_subgraph5/6/7, science Context bands: lt16k, 16k-32k, 32k-48k, 48k-64k

data_source	train	validation
`literesearcher_rag_direct_1-7_16k-32k`	819	44
`literesearcher_rag_direct_1-7_32k-48k`	84	4
`literesearcher_rag_direct_1-7_48k-64k`	14	1
`literesearcher_rag_direct_1-7_lt16k`	4222	120
`literesearcher_rag_direct_wiki_1-7_16k-32k`	1408	40
`literesearcher_rag_direct_wiki_1-7_32k-48k`	105	9
`literesearcher_rag_direct_wiki_1-7_48k-64k`	2	0
`literesearcher_rag_direct_wiki_1-7_lt16k`	4102	127
`literesearcher_rag_mqa_1-7_16k-32k`	429	13
`literesearcher_rag_mqa_1-7_32k-48k`	84	4
`literesearcher_rag_mqa_1-7_48k-64k`	39	1
`literesearcher_rag_mqa_subgraph5_1-7_16k-32k`	346	8
`literesearcher_rag_mqa_subgraph5_1-7_32k-48k`	29	0
`literesearcher_rag_mqa_subgraph5_1-7_48k-64k`	3	0
`literesearcher_rag_mqa_subgraph5_1-7_lt16k`	578	23
`literesearcher_rag_mqa_subgraph6_1-7_16k-32k`	302	12
`literesearcher_rag_mqa_subgraph6_1-7_32k-48k`	15	1
`literesearcher_rag_mqa_subgraph6_1-7_lt16k`	872	30
`literesearcher_rag_mqa_subgraph7_1-7_16k-32k`	116	1
`literesearcher_rag_mqa_subgraph7_1-7_32k-48k`	6	0
`literesearcher_rag_mqa_subgraph7_1-7_lt16k`	366	20
`literesearcher_rag_science_1-7_16k-32k`	399	10
`literesearcher_rag_science_1-7_32k-48k`	59	1
`literesearcher_rag_science_1-7_48k-64k`	18	0
`literesearcher_rag_science_1-7_lt16k`	1270	43

The distribution is heavily skewed toward short contexts: lt16k is ~71% of the validation split, while six buckets have 0-1 validation rows. Per-bucket accuracy on those buckets is not meaningful on its own — weight by row count before drawing conclusions about long-context or multi-hop performance.

Intended use

Stage-2 GRPO training of a tool-using search agent against a local retrieval stack (Milvus + BGE-M3 for search, a page corpus for browse). Upstream notes that Stage 2 continues from a Stage-1 checkpoint; starting directly from an SFT checkpoint skips that warmup.

License

Apache-2.0, inherited from upstream. Please cite `simplex-ai-inc/LiteResearcher-Data`.