SkillRL SFT AlfWorld Prompt Completion: 7,486 Agent Trajectories

7,486 prompt-completion pairs for fine-tuning language models, derived from the SkillRL paper on evolving agents via recursive skill-augmented reinforcement learning. The dataset contains 237 distinct tasks and was authored by yananchen, with a last recorded update in April 2026. All rows contain actions within the admissible action list, indicating a focus on valid agent behavior.

Use Cases

• Fine-tuning language models for agent control based on the described AlfWorld task completions.
• Training models on recursive skill augmentation based on the methodology from the SkillRL paper.
• Benchmarking language model performance on sequential decision-making tasks based on the 237 distinct parsed tasks.
• Studying action validity in agent trajectories based on the 100% match between output actions and the admissible action list.

Strengths

• All 7,486 rows contain actions that exactly match the admissible action list, indicating high internal consistency.
• The dataset is structured around 237 distinct tasks, providing variety in the training scenarios.
• Contiguous runs of the same task have a mean length of 15.0, suggesting sequences of related actions for skill learning.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.

Provenance

Source

Paper: SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning.

Collection Method

Likely generated from agent trajectories in the AlfWorld simulation environment.

Freshness

Last updated 2026-04-10 01:13:19; freshness should be verified.