OpenJev : Open-Sourced Jev AI
Read the Portuguese translation →
Author: Data Science in your pocket · Watch on YouTube
Video authored by the channel shown, not by IANIA; all rights belong to the creator. Transcript and summary generated automatically; the original-language content is preserved.
Original language: English
Original video: OpenJev : Open-Sourced Jev AI · Channel: Data Science in your pocket · 01:33 · Watch on YouTube
Topics: AI agents, AI models, RAG, search and information retrieval
Summary
Section titled “Summary”OpenJev is a community, Jev-inspired adaptation of Qwen 3.5 designed to compare choices rather than generate answers. It classifies a premise and hypothesis as contradiction, entailment, or neutral, then uses entailment scores to rank supplied choices. Its scores can support retrieval, ranking, or a game controller, but OpenJev is not a complete autonomous player and its selections are not guaranteed to be correct.
Key points
Section titled “Key points”- OpenJev is described as a community Jev-inspired adaptation of Qwen 3.5.
- It is trained for classification and reads a premise and hypothesis together.
- A classification head turns the final token representation into contradiction, entailment, and neutral scores.
- Neutral means unresolved, not false.
- The surrounding software supplies choices, and OpenJev scores each choice against the context.
- The adapter keeps entailment scores, takes each choice’s maximum across context windows, and divides by their sum.
- The resulting selection signal can be used for retrieval, ranking, or a game controller.
- A controller supplies state and executes actions; OpenJev supplies scores and is not a complete autonomous player.
- The recommended 4 billion parameter version is V5 for type decisions, but confident selections can still be wrong.
- OpenJev is not the official Jev and does not guarantee correctness.
Chapters
Section titled “Chapters”| Start | Chapter | Summary |
|---|---|---|
| 00:00 | OpenJev and its classification approach | OpenJev adapts Qwen 3.5 for comparing a premise and hypothesis, classifying them as contradiction, entailment, or neutral. |
| 00:31 | Scoring choices | Examples clarify the three classifications, and the adapter turns entailment scores for supplied choices into a selection signal. |
| 01:01 | Uses and limitations | OpenJev can support retrieval, ranking, or game controllers, but it is not autonomous and its predictions may be wrong. |
Quotes
Section titled “Quotes”What if a language model compared choices instead of writing another answer? — 00:00
Neutral does not mean false. — 00:31
It is not a complete autonomous player. — 01:01
Think Qwen reshaped for comparison, not the official Jev and not a guarantee of correctness. — 01:33
Concepts and entities
Section titled “Concepts and entities”OpenJev, Qwen3.5, classification head, entailment, context windows
Full transcript (original language)
Section titled “Full transcript (original language)”00:00 What if a language model compared choices instead of writing another answer? Meet OpenJev, a community Jev-inspired adaptation of Qwen 3.5. Trained for classification, its backbone reads a premise and hypothesis together. A classification head turns the final token’s representation into three scores. Contradiction, entailment, and neutral.
00:31 Imagine this premise: someone is playing guitar music in a room. Someone is making music is supported. The room is silent conflicts. It is night time remains unresolved. Neutral does not mean false. For a decision, the surrounding software supplies possible choices. OpenJev scores each against the supplied context. The adapter keeps entailment scores, takes each choice’s maximum across
01:01 context windows, then divides by their sum. This creates a selection signal for retrieval, ranking, or a game controller. The controller supplies state and executes actions. OpenJev supplies scores. It is not a complete autonomous player. The recommended 4 billion parameter version is V5 for type decisions. But these scores do not guarantee truth. A confident selection can still be
01:33 wrong. Think Qwen reshaped for comparison, not the official Jev and not a guarantee of correctness.