Naluri-561M-jev-distilled

Install torch, transformers>=4.48, safetensors, and huggingface_hub, then download the bundled naluri.py from the model repo and call naluri.from_pretrained with device set to cpu or cuda. The standard transformers pipeline and the hosted inference widget cannot run it because its decision head is custom code.

In one additional training epoch, about 23% of the examples are black-box distilled from Jev: new questions about texts Naluri had already seen, labelled with Jev's probabilities through Jev's public API. The remaining examples are Naluri's own gold-labelled data.

On System One Bench, Naluri-561M-encoder scores 64.9 and Jev 1.14 from the TypeSafe API scores 79.3. The distilled model's own score is listed as pending. The model card states it does not reach Jev's accuracy and is not a drop-in replacement for Jev.

The repository is licensed cc-by-nc-4.0, a non-commercial license. The model is public and not gated, but the non-commercial terms apply to use of the weights.

It uses the same input format and the same three question types as the 561M model: noul, choice, and score. It is multilingual across 18 languages, including English, Indonesian, Spanish, French, German, Portuguese, Chinese, Japanese, Korean, Arabic, Hindi, Russian, Vietnamese, Thai, Turkish, Italian, and Swahili.

The model card describes the experiment as learning from a commercial decision model's outputs alone for about $1.35 of API credit, spent labelling new questions with Jev's probabilities through its public API.

The distillation script distill.py and a full report are published in the naluri GitHub repository at github.com/jerichosiahaya/naluri, alongside the code for the Naluri model family.