NLI as a lightweight, general-purpose primitive: openjev operationalizes the idea that a single entailment classifier can serve many downstream roles — reranking, grading against a reference, safety checks, and even selecting actions in games — by fine-tuning Qwen3.5 as a three-way Natural Language Inference model and exposing task-specific latent heads.
What Sets It Apart
- Single NLI interface for many tasks — so what? You can turn question-answering, multiple-choice ranking, and reference-based grading into the same entailment decision, simplifying toolchains and enabling zero-shot use without per-task finetuning.
- Frozen-latent + tiny MLP heads for multiple-choice — so what? The heavy backbone can be frozen and kept shared while small per-task heads (soft-BCE training) give competitive option scoring with cheap head training and fast iteration.
- Hands-on utilities and demos — so what? The repo ships a ready Qwen3.5-4B NLI checkpoint, a Python helper (OpenJevCrossEncoder) for predict/rerank/latents, MLP head code, and demos (Flappy Bird, Doom) showing text- and pixel-based zero-shot control via the vision tower.
- Empirical NLI performance — so what? The 4B variant reports strong MNLI-style accuracy (~0.90), indicating it is a reliable entailment scorer for downstream reranking and grading tasks.
Who it's for and trade-offs
Great fit if you need a compact way to convert many NLP tasks into entailment decisions: researchers or engineers who want an NLI reranker, a grading/QA verifier, or a low-cost per-task head workflow. The project is practical for prototyping and for setups that prefer freezing a large backbone and training small heads.
Look elsewhere if you need turnkey task-specific SOTA models or an out-of-the-box multimodal policy system: using the vision tower for pixel inputs is demonstrated but not the primary focus, and heavy customization or full single-task finetuning may outperform the NLI-as-primitive approach in narrow benchmarks.
Where it fits
Openjev sits between generic pretrained LLMs and task-specific classifiers: it repurposes a decoder-only Qwen3.5 into a cross-encoder NLI model, making it a useful reranker/validator layer in pipelines (retrieval+rerank, answer verification, safety filters) and a bridge to small per-task heads for efficient fine-tuning.
