README.md

---
pretty_name: Simulated Scholar Search (S3) RL Dataset
license: apache-2.0
language:
  - en
task_categories:
  - question-answering
size_categories:
  - 10K<n<100K
tags:
  - agentic-rl
  - deep-research
  - literature-search
  - reinforcement-learning
  - scientific-literature
  - simscholar
  - s2orc
  - synthetic-data
configs:
  - config_name: default
    data_files:
      - split: train
        path: trainer_13k.jsonl
---

<p align="center">
  <img src="https://raw.githubusercontent.com/trillion-labs/scholar-search-rl/main/assets/s3-logo.png" alt="Simulated Scholar Search (S3)" width="160">
</p>

<h1 align="center">S3 RL Dataset</h1>

<p align="center">
  Synthetic literature-search questions for agentic reinforcement learning.
</p>

<p align="center">
  <a href="https://github.com/trillion-labs/scholar-search-rl">Code</a> ·
  <a href="https://huggingface.co/collections/trillionlabs/simulated-scholar-search-s3-models-and-datasets-6a42188a4d64a10aa86ccd61">S3 collection</a> ·
  <a href="https://huggingface.co/datasets/AlgorithmicResearchGroup/s2orc-cs-enriched">Source corpus</a>
</p>

The S3 RL dataset contains 13,000 synthetic, single-hop questions for training
and analyzing scientific-literature search agents. Each question is grounded
in one paper from a fixed corpus of approximately 1.12 million
computer-science papers.

## At a glance

| Property | Value |
| --- | --- |
| Rows | 13,000 |
| Language | English |
| Task | Tool-assisted scientific-literature QA |
| Scope | Single-hop; one source paper per question |
| Format | JSON Lines |

### Composition

| Answer type | Rows |
| --- | ---: |
| Abstract | 5,000 |
| Value | 5,000 |
| Identity | 3,000 |

| Query style | Rows |
| --- | ---: |
| Content conjunction | 3,000 |
| Context conjunction | 3,000 |
| Detail cue | 3,000 |
| Named | 2,000 |
| Paraphrastic | 2,000 |

## Schema

| Field | Type | Description |
| --- | --- | --- |
| `sample_id` | string | Stable generated-item identifier |
| `query` | string | Question presented to the agent |
| `gold_answer` | string | Reference answer used for grading |
| `anchor` | string | Query framing used during synthesis |
| `answer_type` | string | `abstract`, `value`, or `identity` |

## Load the dataset

```python
from datasets import load_dataset

dataset = load_dataset("trillionlabs/SimScholar-RL", split="train")
print(dataset[0])
```

## Research use

The dataset is intended for research on literature-search agents,
retrieval-grounded QA, tool-use policies, curriculum design, and agentic
reinforcement learning.

## Limitations

- Questions and reference answers are synthetic and may contain generation or
  grading errors.
- The source corpus covers computer science rather than the full scientific
  literature.
- A fixed local retrieval environment differs from live scholarly search.
- Some questions may be answerable from model memory without tool use.
- Verbatim overlap with source-paper text has not been exhaustively audited.

## License

This dataset is licensed under the
[Apache License 2.0](https://www.apache.org/licenses/LICENSE-2.0).
Third-party paper text, excerpts, and metadata retain their original rights and
terms; the Apache license does not supersede them.