Anarkali 150M

Anarkali 150M is an ONNX release of a packed encoder choice model for structured decision-support questions. It ranks the candidate options supplied in an input request; it does not establish real-world correctness and must not be used as an autonomous decision-maker.

Files

• model.onnx: validated FP32 inference graph.
• model.optimized.onnx: ORT-optimized FP32 graph.
• tokenizer.json: tokenizer required by the Anarkali runtime.
• anarkali.json: runtime configuration and provenance.

Validation

The release passed PyTorch-to-ONNX parity on 200 held-out-format development rows:

• Argmax mismatches: 0
• Maximum probability drift: 1.96e-6
• FP32 CPU latency (batch 1, tokenization and packing included): p50 446 ms, p95 704 ms.

These metrics measure agreement with evaluation labels/teachers, not proof of real-world correctness or safety. The model must be reviewed against the deployment's baseline, acceptance criteria, and human oversight requirements.

Usage

Install the project runtime and load the Hub repository:

from anarkali.engine import Engine

engine = Engine.load("toufiqqureshi651/anarkali")

See the project repository for request schemas, evaluation scripts, and service integration guidance.

License and provenance

Apache-2.0. The model uses the Ettin 150M encoder and was trained on the anarkali-decisions-v3 data revision recorded in anarkali.json. Refer to the project LICENSE, NOTICE, and documentation for third-party licenses, data lineage, and limitations.