You need to agree to share your contact information to access this model

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

YAML Metadata Warning:empty or missing yaml metadata in repo card

Check out the documentation for more information.

SurrealML .surml Output Normaliser Visibility Gap — PoC

Demonstrates that SurMlFile.load() and raw_compute() in surrealml v0.0.4 do not surface the output normaliser configuration embedded in a .surml header. A .surml file with an output normaliser returns different numeric values through raw_compute() and buffered_compute(), with no API-level warning from either path.

Setup

pip install surrealml==0.0.4 onnx==1.21.0 "onnxruntime>=1.20.0" "ml_dtypes>=0.3.0"

aarch64 note: surrealml v0.0.4 ships an x86-64 ONNX Runtime binary inside libc_wrapper.so. On aarch64 systems, build and use the included shim to redirect ONNX Runtime loading to the system-installed aarch64 library: bash build_dlopen_shim.sh then prefix commands with LD_PRELOAD=./dlopen_intercept.so. This is an environment workaround and is not part of the reported issue.

Step 1: Create .surml with embedded output normaliser

python3 create_model.py .

Output:

ONNX_MODEL_VALID=True
SURML_SAVED=True
OUTPUT_NORMALISER_TYPE=z_score
OUTPUT_NORMALISER_P1=0.0
OUTPUT_NORMALISER_P2=999.0

Step 2: Inspect via load() and raw_compute — normaliser config not surfaced, no warning

python3 inspect_model.py .

Output:

SURML_LOADED=True
LOAD_SURFACES_NORMALISER_CONFIG=False
RAW_COMPUTE_OUTPUT=1.0
WARNING_EMITTED=False

SurMlFile.load() returns only (file_id, name, description, version). The output normaliser type, parameters, and mapping are not included. raw_compute() returns the raw ONNX output with no warning that an output normaliser exists.

Step 3: Runtime — buffered_compute applies output normaliser, result diverges

python3 reproduce.py .

Output:

RAW_VALUE=1.0
BUFFERED_VALUE=999.0
OUTPUT_FLIP_CONFIRMED=true
WARNING_EMITTED=False

Summary

	Output	Normaliser applied	Warning
`raw_compute([1.0])`	1.0	No	None
`buffered_compute({"x": 1.0})`	999.0	Yes (z_score)	None

The gap is that neither load() nor raw_compute() surfaces or warns about the output normaliser. Workflows that use raw_compute() as a validation path before using buffered_compute() need a reliable API signal that the buffered path will apply an output transformation. Without that signal, the validation result can differ from the served result without the caller being alerted.

Files

File	SHA-256
`create_model.py`	`381ed9f669fdb138a3ec98debd49c3cbdb31a8f59993b017249fbd1d12271b51`
`inspect_model.py`	`ad8e648ce5a3750b39e0729e2a8e3f3c3e45b98236b377ce12b737073090c3ab`
`reproduce.py`	`dd5db3e7f802b8cb8f45fa6beb306a7eaaa9faece58949545526b1601732ef31`
`dlopen_intercept.c`	`2eea3ba286484c22dcebc768df3a0cf756224f496aef993a05c74afbef00e29b`
`expected_output.txt`	`ba54c3e051270feee4ba28fab7cfc8a7069a3738d894751bfd08a283a24bc929`
`model.surml`	`7bf6d816857844e41b2e727b6aad8871cff760debfd84350f5dfc59b105f3e57`
`model.onnx`	`7bd0c40611de64286d1eab80f116f67addf0ec221f59b37c478b508ade25c9a8`

Downloads last month: -; Downloads are not tracked for this model. How to track

Inference Providers NEW

This model isn't deployed by any Inference Provider. 🙋 Ask for provider support