fix(sparse): preserve small positive SPLADE weights with log1p

[RAMZI0TO99]

For float32 logits [[[1e-8, 0.0, -1.0, 1.0]]] with attention mask [[1]], the current log(1+x) postprocessing drops dimension 0 because adding the small value to 1 rounds to 1. The resulting sparse indices are [3] instead of [0, 3].

This uses np.log1p to evaluate the same mathematical log-ReLU formula accurately for small positives. Masking, maximum pooling, and nonzero filtering are unchanged. The tests compare tiny values against an independent scalar reference and cover negatives/zeros, padding exclusion, batch order, pooling, empty output, input preservation, and missing-mask behavior. This is a numerical-accuracy fix; retrieval-quality improvement is not claimed.

Validation on Windows 11, Python 3.13.5, NumPy 2.3.5, FastEmbed 0.8.1:

• 18 selected tests passed: python -m pytest -q tests/test_splade_log1p.py.
• Local diff-scoped AST docstring coverage: 9/9 functions (100%, above the reported 80% threshold). CodeRabbit's own result is separate.
• Ruff 0.3.4 lint/format and staged whitespace checks passed.
• Repository mypy and pyright checks passed on Python 3.13.5.
• Six existing actual-model SPLADE checks passed on this split checkout using revision 600ce23fd7611e71694fcf724f46778f1ad24c1c. Canonical expectations were unchanged. The single-embedding test used its existing model-selection hook to select SPLADE.
• Model test logs retain one non-failing anyio assertion-rewrite warning.

The full model/OS/Python CI matrix was not run locally.

Split from #766 following the maintainer's request for one PR per problem. This branch is based directly on current main and contains only this problem's production change and regression module.

[coderabbitai]

Navigate logical layers of code changes, visualize relationships, and explore their blast radius.

No actionable comments were generated in the recent review. 🎉

ℹ️ Recent review info

⚙️ Run configuration

• Configuration used: Organization UI
• Review profile: CHILL
• Plan: Advanced
• Run ID: ed75dc1c-8400-470d-abfe-50a3ccfc18a3

📥 Commits

Reviewing files that changed from the base of the PR and between ebd89a0 and 2c72313.

📒 Files selected for processing (1)

• fastembed/sparse/splade_pp.py

💤 Files with no reviewable changes (1)

• fastembed/sparse/splade_pp.py

Included review availability: This review used your included allowance. Your plan provides up to 8 included reviews per hour; 7 remain after this review.

---

📝 Walkthrough

Walkthrough

SpladePP._post_process_onnx_output now uses np.log1p for the log transform. Attention-mask weighting and subsequent scoring are unchanged.

Priority: ⬇️ Low

Estimated code review effort: 1 (Trivial) | ~2 minutes

Change: Bug fix

Merge Risk: ⚪ Minimal · up to 2c723

This change preserves very small positive SPLADE weights that were previously rounded away. It carries little risk and is ready to merge after the normal CI checks.

🚥 Pre-merge checks | ✅ 5

✅ Passed checks (5 passed)

Check name	Status	Explanation
Title check	✅ Passed	The title clearly and concisely identifies the SPLADE numerical-accuracy fix using log1p.
Description check	✅ Passed	The description explains the small-positive-weight issue, the log1p fix, and the reported validation.
Docstring Coverage	✅ Passed	Docstring coverage is 88.89% which is sufficient. The required threshold is 80.00%. Docstring coverage is scoped to functions touched by this diff. Analyzed 9 functions across 2 files.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create a new PR

• Autopilot · Keep fixing CodeRabbit findings and required CI, and resolving merge conflicts

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[joein]

Thanks for the PR! I ran it on ~9.5k texts (SciFact, Quora, AG News) and saw no measurable difference: no tokens gained or lost, and SciFact nDCG@10 was identical. Since sentence-transformers also uses log1p, I'm ok with merging it, but I dropped the new test file and the docstring to keep the change to the one-liner.