README
openjev
A Qwen3.5 decoder fine-tuned as an NLI cross-encoder. One forward pass over
Premise / Hypothesis returns contradiction · entailment · neutral, and the premise can be text
or an image — so the same 0.8B model does RAG fact-checking, zero-shot reranking, VQA-style
image judging and frame-by-frame video scanning.
Served checkpoint: AlexWortega/openjev @
qwen3.5-0.8b-nli-v2s-long — 0.8B, 4k context, stage 3 of the v2s recipe (NLI mixture +
faithfulness / instruction-following / false-premise data, then long documents).
| tab | what it does |
|---|---|
| 🧩 NLI | premise + hypothesis → 3-class probabilities |
| 🔎 RAG fact-check | document + one claim per line → supported / contradicted / unsupported |
| 🥇 Rerank | question + options → ranked by entailment |
| 🖼️ Image NLI | the picture is the premise |
| 🎬 Video scan | sample frames, track one hypothesis over time |
| 📊 About | benchmarks and the agent videos |
openjev_core.py holds the model and the scoring function (no gradio imports, so it can be run
against a local checkpoint with OPENJEV_MODEL=/path/to/ckpt OPENJEV_SUBFOLDER=); app.py is the UI.
Similar repos
README
openjev
A Qwen3.5 decoder fine-tuned as an NLI cross-encoder. One forward pass over
Premise / Hypothesis returns contradiction · entailment · neutral, and the premise can be text
or an image — so the same 0.8B model does RAG fact-checking, zero-shot reranking, VQA-style
image judging and frame-by-frame video scanning.
Served checkpoint: AlexWortega/openjev @
qwen3.5-0.8b-nli-v2s-long — 0.8B, 4k context, stage 3 of the v2s recipe (NLI mixture +
faithfulness / instruction-following / false-premise data, then long documents).
| tab | what it does |
|---|---|
| 🧩 NLI | premise + hypothesis → 3-class probabilities |
| 🔎 RAG fact-check | document + one claim per line → supported / contradicted / unsupported |
| 🥇 Rerank | question + options → ranked by entailment |
| 🖼️ Image NLI | the picture is the premise |
| 🎬 Video scan | sample frames, track one hypothesis over time |
| 📊 About | benchmarks and the agent videos |
openjev_core.py holds the model and the scoring function (no gradio imports, so it can be run
against a local checkpoint with OPENJEV_MODEL=/path/to/ckpt OPENJEV_SUBFOLDER=); app.py is the UI.