kevinqz/Pulpie-Orange-Base-CoreAI

Canonical: kevinqz/Pulpie-Orange-Base-CoreAI — source of truth.

Pulpie Orange Base (fabric)

An Apple Core AI conversion of feyninc/pulpie-orange-base — a token classifier (bidirectional encoder) that maps (input_ids, attention_mask) to per-token logits. Produced by coreai-fabric and indexed by coreai-catalog.

Encoder, not a chat model. This is a single-forward classifier — no text generation, no KV-cache. The host owns the tokenizer (use the upstream tokenizer at feyninc/pulpie-orange-base), feeds token ids, and reads the per-token argmax.

Model facts

Field	Value
Parameters	0.6B
Architecture	transformer
Capabilities	token-classification
Labels	2
Sequence length	64 (static)
Quantization / precision	none / float32
On-disk size	2.3 GB
Asset kind	single-graph encoder ((input_ids, attention_mask) -> per-token logits)
assetVersion	2.0

Use it — this needs host code you supply

The bundle is a single static-sequence graph: (input_ids, attention_mask) in → per-token logits out. You supply the upstream tokenizer and the argmax / label mapping in your host code (Swift or Python). Token ids are int32 at the graph boundary; pad or truncate to the static sequence length (see Model facts).

pip install coreai-catalog && coreai-catalog install pulpie-orange-base

Requirements

• Deployment: macOS 27.0+ / iOS 27.0+, Xcode 27+. The asset serializes with minimum_os v27, so the on-device Swift runtime requires macOS/iOS 27+. A Mac on macOS 26 can convert and inspect it but not run it on-device.
• Apple Silicon.

Verification (output parity)

• Gate A (structure): passed — the bundle's layout + metadata were validated; the graph loads.
• Gate B — graph_output_cosine: 1.000000 min output cosine (median 1.000000) vs the fp32 torch reference over 8 seeded (input_ids, attention_mask), measured on apple_silicon. The encoder analog of the LLM logit-parity: it certifies the export computes the SAME per-token logits as the source — a conversion-fidelity metric, not task accuracy.
• This certifies the export is numerically faithful to the source encoder — it does NOT certify downstream task accuracy on your data. Reproduce with coreai-fabric verify.

Provenance

Field	Value
Base model	feyninc/pulpie-orange-base @ d8f4eaabc8647b96034b906ab3864572e1489c10
Converted by	models/eurobert/export.py (version not reported)
Recipe	pulpie-orange-base (recipe_source: fabric)
Precision / quantization	float32 / none
Conversion date	2026-07-07

Machine-readable, in this repo: parity-report.json · reproduce-manifest.json · LICENSE.

License and attribution

Weights licensed apache-2.0 — see the bundled LICENSE. This artifact is a converted derivative of the base model: its weights were converted to Apple Core AI format. The conversion itself is community work.

Links

• Base model: feyninc/pulpie-orange-base
• Reproduce: recipe pulpie-orange-base
• Index: coreai-catalog
• HF Collection

The on-device Core AI ecosystem

• coreai-fabric — the reproducible recipe → .aimodel pipeline that produced this asset.
• coreai-catalog — the index of Core AI models with provenance and integration snippets.
• apple/coreai-models — Apple's official exporters and runtimes.

Not affiliated with Apple

Community conversion. Not produced, hosted, or endorsed by Apple. Apple and Core AI are trademarks of Apple Inc., used here only to describe the target runtime/format.