Update orphan GRB classifier model retrained with sklearn 1.5.2

[sedlachevre]

IMPORTANT: Please create an issue first before opening a Pull Request.
Linked to issue(s): Closes #683 (comment)

If this is a new release, did you issue the corresponding schema in fink-client?

N/A

What changes were proposed in this pull request?

The orphan GRB afterglow classifier model was retrained with scikit-learn 1.5.2 to replace the old model which was incompatible with the current sklearn versions due to the removal of sklearn.ensemble._gb_losses in sklearn 1.0.

How is the issue this PR is referenced against solved with this PR?

How was this patch tested?

The new model was locally tested on a sample of 1915 simulated orphan afterglow light curves and 695195 events from the ELAsTiCC data set (playing the role of background) and showed great accuracy.

[JulienPeloton]

Thank you @sedlachevre ! Could you uncomment this line in your fork:

fink-science/fink_science/rubin/orphans/processor.py

Line 114 in 7c7bfbf

# spark_unit_tests(globs)

and push, to trigger the test suite on this module?

[bregeon]

@JulienPeloton I am a bit surprised by the error of the test-suite above:
/opt/fink-broker/spark/bin/load-spark-env.sh: line 68: ps: command not found
indeed when I run the container locally, I have the same error and the procps package is indeed not installed in the container... it does not appear to be mandatory, but well :)

@sedlachevre, actually, the error above is not a blocker, the real issue is that we need to add the following import from fink_science.tester import spark_unit_tests in processor.py (see rubin/hostless_detection/processor.py)

[JulienPeloton]

Yes, ps is not installed in the container, but we do not need it (despite Spark complaining, it is not required for this work).

[JulienPeloton]

Thanks for the change -- looking at the continuous integration, the test on orphans is failing usinng previous setup:

test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3938135Z File "/__w/fink-science/fink-science/fink_science/__init__.py", line ?, in __main__.orphan_grb
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3938873Z Failed example:
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3939206Z     df.filter(df["prob"] > 0.0).count()
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3941283Z Expected:
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3941904Z     0
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3942164Z Got:
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3942409Z     6
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3942720Z **********************************************************************
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3943189Z 1 items had failures:
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3943533Z    1 of  10 in __main__.orphan_grb
test-suite (rubin)	UNKNOWN STEP	2026-04-03T11:50:55.3943921Z ***Test Failed*** 1 failures.

Is that expected? Can you inspect these 6 objects that have non-zero probability?

[sedlachevre]

I ran the test locally on the same data and I also got 6 objects with a non-zero probability but these probabilities are very small (p<0.00003). This is expected with our machine learning model and should not be a problem for orphan detection. Is it possible to only count objects with a probability higher than 0.1 instead of 0.0 (df.filter(df["prob"] > 0.1).count()) ?

[JulienPeloton]

Sure! You can update the test accordingly.