codesoda/kev-heads

Kev heads (pickle-free safetensors)

Converted pointer heads for Kev decision models, for runtimes that must not execute PyTorch pickle files.

Upstream Kev ships each head as head.pt, a PyTorch pickle. This repository holds the same tensors as head.safetensors, plus head.meta.json with the temperature and metadata. kev-rs reads only these files. SystemOne s1 setup downloads them together with the pinned base model and LoRA adapter.

Nothing here is retrained or changed. The conversion copies tensors and metadata only.

Files

Checkpoint	Source adapter (with head.pt)	Source head.pt SHA-256	Base model
kev-0.6b	jaredpalmer/kev-0.6b @ dece6dba8d43	ce6cd9ffc54db41c…	Qwen/Qwen3-0.6B-Base @ da87bfb608c1
kev-0.8b	jaredpalmer/kev-0.8b @ 54f4f8777356	39f4343ccccc65e5…	Qwen/Qwen3.5-0.8B-Base @ dc7cdfe2ee41
kev-4b	jaredpalmer/kev-4b @ 485ace870359	d8f796da36ff7bd7…	Qwen/Qwen3.5-4B-Base @ 1001bb4d826a

Each directory holds head.safetensors and head.meta.json. SHA256SUMS lists every file.

Provenance

• Upstream: jaredpalmer/kev @ 557598fced1dada75dfbf36ed144dce309ac6ceb (Apache-2.0).
• Converter: kev-convert-head in kev-rs benchmarks/baseline, kev-rs v0.1.1. The expected output checksums are recorded in kev-rs manifests/sources.json.
• kev-rs checks its Rust runtime against frozen upstream PyTorch and MLX outputs that use these heads.

License

Apache-2.0, the same as upstream Kev, its adapters and the Qwen3 / Qwen3.5 base models. Not affiliated with Kev's authors or with TypeSafe AI.