OSU-NLP-GroupOSU-NLP-GroupOfficialยท1 Agent Skills Included

EarlyExperience

Reward-free training data generation for interactive environments

Generates expert, world-modeling, and self-reflection training data for interactive environments without reward signals. Eliminates the need for costly reinforcement learning rollouts and manual data labeling across ten benchmark environments. Guides coding agents through smoke tests, sampling decisions, and output formatting to produce ready-to-train datasets faster.
npx skills add OSU-NLP-Group/EarlyExperience --all -g -y

All Skills in This Repository (1)

Pure Emerald Level Indicators

Frequently Asked Questions

FAQPage Schema
How to install EarlyExperience?โ–ผ

Run `npx skills add OSU-NLP-Group/EarlyExperience --all -g -y` in your terminal to install the skill globally.

What is early experience learning?โ–ผ

It is a training method between imitation and reinforcement learning where a model learns from the outcomes of its own alternative actions, without needing reward signals.

Which environments does EarlyExperience support?โ–ผ

It ships notes and data for ten environments including ALFWorld, WebShop, BFCL, ScienceWorld, TravelPlanner, AppWorld, and TextCraft.

Does EarlyExperience work with any coding agent?โ–ผ

Yes. The skill folder follows the standard SKILL.md format, so any compatible agent can read it and follow the data-generation workflow.

Where can I get the pre-generated training data?โ–ผ

All expert, world-modeling, and reflection datasets are published on Hugging Face under the osunlp/early-experience dataset.

Related Repositories in Education & Research

View All in Education & Researchโ†’