SkillRL-SFT-Data: Instruction-Output Pairs for Agent Training

Instruction-output pairs for training base agent policies in interactive decision-making environments. The dataset was created by author Jianwen for the SkillRL research paper and was last updated on 2026-04-12. It contains examples with structured instructions and corresponding expert actions.

Use Cases

• Supervised fine-tuning of language models for agent control based on the provided instruction-output pairs.
• Training base policies for agents operating in ALFWorld, WebShop, and Search environments.
• Research on hierarchical skill learning and skill transfer using the retrieved skill context from the SkillBank.
• Benchmarking agent performance on instruction-following tasks in interactive decision-making settings.

Strengths

• Dataset is explicitly designed for and used in a published research paper (SkillRL).
• Covers three distinct interactive decision-making environments: ALFWorld, WebShop, and Search.
• Each example contains structured instruction with retrieved skill context and expert action output.

Limitations

• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.
• Description metadata is limited; actual data quality requires manual inspection after download.

Provenance

Source

huggingface

Collection Method

Created for the paper 'SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning'.

Freshness

Last updated 2026-04-12 04:51:35; freshness should be verified.