AI BriefingPickopenjev: From Game Play to RAG with a Single NLI Model
AlexWortega/openjev
About the project
A single cross-encoder that takes premises and hypotheses as input and answers with one of three labels: true, false, or neutral. Using only this simple NLI principle, it performs various tasks such as search result reranking, answer grading, and content filtering.
Given a game state and related sentences, it selects the action with the highest entailment probability. It enables zero-shot play of Flappy Bird and Doom using both text states and pixel inputs.
Trained on Qwen3.5-4B, it also offers a 35B MoE variant with a frozen backbone and a small MLP head. It is designed for general-purpose use without task-specific training.
It can be easily loaded via the Transformers library and provides core methods such as predict, rerank, and grade. Under the MIT license, it facilitates experimentation and integration within the open-source ecosystem.
AlexWortega/openjev
The original page has no description.
text-classification
This introduction was generated automatically by AI. Check the original for the author's claims and context. Copyright belongs to the original author.
Our guide explains how the AI works. Report errors, attribution issues, or removal requests via Contact.