Skip to content

MODEL_PROFILE.md

Latest commit

 

History

History
118 lines (81 loc) · 6.07 KB

MODEL_PROFILE.md

File metadata and controls

118 lines (81 loc) · 6.07 KB
Raw

Local specialist model profiles

Models are optional evidence providers. Polymorph does not require a generative model and does not send record payload values to a model interface.

Research-only descriptor encoder

Comparison profile:

sentence-transformers/paraphrase-multilingual-MiniLM-L12-v2

Role: retrieve plausible target fields from schema descriptor text such as names, aliases, descriptions and container names.

Selected properties:

  • 384-dimensional embeddings
  • multilingual descriptor support
  • ONNX Runtime on CPU
  • native SentencePiece tokenization to keep the 250,000-piece vocabulary compact in memory
  • quantized x86_64/AVX2 and ARM64 artifacts
  • Apache License 2.0 upstream
  • no generative decoding path in Polymorph

Pinned upstream revision:

e8f8c211226b894fcb81acc59f3b34ba3efd5f42

The Hugging Face weight card declares Apache-2.0. Sentence Transformers documents the English teacher model as trained on several datasets including msmarco-triplets, then documents this profile as its multilingual distilled version. Microsoft's MS MARCO terms restrict that dataset to noncommercial research and explicitly recommend an independent legal review for research outputs used in products. Polymorph therefore does not install or enable this profile by default. Keep it out of a commercial path unless a proper rights review clears it. Explicit lab installation uses bootstrap's --include-research-encoder; activation still requires --models or --model-dir.

The 42-case local safety regression completed in 0.926 seconds at 48.59 fields per second with 284.52 MiB peak RSS. It made the same 17 safe automatic decisions and the same 31 correct suggestions as the deterministic and two-model profiles.

Experimental ambiguity reranker

Research-only comparison profile:

cross-encoder/mmarco-mMiniLMv2-L12-H384-v1

Role: compare only the top few target candidates when initial deterministic/embedding evidence is close. The reranker is not called for clear cases.

This profile is not installed by default and --models does not enable it. The Hugging Face weight card declares Apache-2.0, but the model card documents mMARCO/MS MARCO training data. Microsoft's official MS MARCO terms restrict that dataset to noncommercial research. That provenance is a bad fit for a planned paid edition even though Polymorph does not redistribute the weights. Keep it out of commercial evaluation unless a proper rights review clears it. Bootstrap's --include-research-reranker switch installs it. Runtime activation still requires --reranker-dir.

Selected properties:

  • multilingual cross-encoder
  • quantized ONNX artifacts around 119 MB
  • AVX2 x86_64 and ARM64 variants
  • Apache License 2.0 upstream
  • no record-value interface

Pinned upstream revision:

1427fd652930e4ba29e8149678df786c240d8825

The reranker score is bounded into evidence used for candidate ordering. It is not treated as a calibrated probability.

Automatic-approval rule

Neither profile is an authorization mechanism. Automatic promotion requires the selected target to also be the independently strongest deterministic target, to exceed the deterministic floor and to have a sufficient deterministic margin.

If model evidence changes the winner away from the deterministic winner, the mapping becomes review-required.

Why not use one larger model everywhere

Larger specialist models can improve retrieval quality, but they increase startup time, memory, download size and the cost of running on ordinary CPUs. They also do not solve missing business context.

Current lab candidates include the Qwen3 0.6B embedding/reranking pair for a future precision mode. They support a much larger multilingual surface but are substantially heavier and the reranker is implemented on a causal-language-model base. They should earn inclusion by improving Polymorph's own source-separated mapping corpus without introducing incorrect automatic promotions.

multilingual-e5-small is another useful benchmark candidate with broad multilingual support. MiniLM remains the compact lab baseline because the upstream profile we pin provides practical quantized AVX2 and ARM64 artifacts for the architectures we want to measure directly. Neither is a default until its full training provenance is commercially reviewable and it improves the held-out Polymorph corpus.

Supply-chain pinning

Profile installation is pinned to exact upstream revisions. Every required ONNX, tokenizer and configuration asset is compared with a built-in SHA-256 value. Polymorph also records those expected hashes in polymorph-model-manifest.json and verifies them again before every load.

Runtime loading verifies the manifest and uses local files directly. Model weights are not redistributed in this repository.

Resource behavior

Profile verification is eager. ONNX sessions and tokenizers are lazy, so a deterministic mapping does not pay model startup or inference cost. The reranker is also skipped whenever deterministic evidence is already decisive.

On the current Windows x86_64 reference machine, Python 3.11 and ONNX Runtime 1.29 produced these local measurements:

  • encoder verification only: about 33 MiB process RSS
  • encoder after first inference: about 240 MiB process RSS
  • 42-case safety regression with both profiles: 75.48 MiB before, 427.29 MiB after, 437.87 MiB peak and 2.325 s wall time
  • both installed profiles after removing obsolete JSON tokenizer copies: about 247 MB on disk
  • research encoder alone on disk: 123.5 MB, or 117.81 MiB

These numbers are evidence for one machine, not a hardware guarantee. Operators should run the included benchmark on their target CPU and corpus. The ONNX graph optimizer, memory pattern and prepacking defaults stay enabled because disabling them changed scores or made private-memory use worse in local comparison runs. Only the encoder CPU memory arena is disabled, saving a modest amount of steady RSS for a small latency cost on the reference machine.

The ARM64 artifacts are pinned and architecture-selected, but the current hosted model smoke runs on x86_64 only. ARM64 remains unverified until a native runner executes installation and inference.