| base_model: C-Tianyu/NanoJev | |
| tags: | |
| - coreml | |
| - decision-model | |
| - conversion-recipe | |
| # NanoJev Core ML conversion source — no weights | |
| This repository contains a validated conversion recipe for the [current NanoJev unified-games checkpoint](https://huggingface.co/C-Tianyu/NanoJev), pinned to revision `047b927b30882a1138fc504821b82ac145a4b81a`. **It does not contain a converted model or trained weights.** The upstream model card labels its source code MIT but does not state a redistribution license for the trained weights. Converted Core ML weights will be published only if that permission is clarified. | |
| The source preserves NanoJev's Qwen3 candidate encoder, trained scalar head, and Choice-specific attention over the full candidate set. It also implements the Boolean and Score output paths. `source/assets.lock.json` records the selected checkpoint SHA256 `f68c47d66998231b86b7e91b4ed5e82ae23acf104c8b7cd6d165c3ac7b7ffe1b`, which contains 596,250,498 trained parameters including heads. The exporter creates separate candidate-encoder and decision-head Core ML packages so one backbone serves all three question types. | |
| The local FP16 L128/K4 export passed native parity on Choice, Boolean, and Score (maximum logit error `1.67e-6`). Core ML selected the same answer as native PyTorch on all three requests; maximum probability errors were `0.003851`, `0.000847`, and `0.005882`, respectively. The encoder package is about 1.1 GiB and the shared head about 420 KiB. The packages remain local because trained-weight redistribution is unresolved. These are parity checks, not the Decision Index or 2048 benchmark. | |
| The current checkpoint is newer than the one likely used for the tracker entry, whose exact revision has not been verified. A local Core ML package from this recipe should not be described as reproducing the tracker's 26.19 score without that identity check and the official suite run. | |
| See `source/README.md` for commands and `source/RESULTS.md` for the local results. This repository's conversion scripts are original Fluid Inference code. NanoJev's source code and checkpoint remain at the upstream repository; the base Qwen model keeps its own license. | |
| `LICENSE-CODE` covers this repository's conversion scripts only. It does not license the upstream trained NanoJev weights. | |