LiteResearcher: Two-Stage Curriculum Prompts for Agentic RL Training

Companion training data for the LiteResearcher paper, which describes a low-cost, scalable Agentic RL training framework for deep-research agents. The dataset contains the two-stage curriculum of question–answer prompts used to train the LiteResearcher-4B model with on-policy GRPO+TIS against a local search/browse environment. Both curriculum stages share the same validation set.

Use Cases

• Training or fine-tuning research-oriented AI agents based on the described two-stage curriculum.
• Studying curriculum learning strategies for reinforcement learning agents based on the described prompt structure.
• Benchmarking agent performance in local search/browse environments based on the provided validation set.
• Reproducing or extending the LiteResearcher training methodology described in the companion paper.

Strengths

• Dataset is explicitly linked to a published research paper (LiteResearcher), providing academic context.
• Describes a structured two-stage curriculum with a shared validation set, suggesting a defined training approach.
• Last updated on 2026-06-05, indicating recent maintenance.

Limitations

• The underlying webpage corpus of ~32 million records is not included in this dataset.
• Column names, sample data, file formats, size, and row counts are unknown.
• License information is unspecified, which may restrict usage.

Provenance

Source

simplex-ai-inc on Hugging Face

Collection Method

Created as companion training data for the LiteResearcher paper; likely generated for the described RL training framework.

Freshness

Last updated 2026-06-05 12:03:18.