fix: upgrade vulnerable dependencies in TORR benchmark requirements

[yoavkatz]

Resolves the Dependabot alerts tracked in #1978.

Where the alerts come from

The repo has only two dependency manifests (no lockfiles, no setup.py, no conda envs) and no .github/dependabot.yml, so Dependabot auto-discovers both:

1. pyproject.toml
2. prepare/benchmarks/torr/requirements.txt

All three alerted packages trace to prepare/benchmarks/torr/requirements.txt — the only manifest that still pinned them below the patched releases. The pyproject.toml occurrences are not the source: the tests extra pins scikit-learn<=1.5.2, a range that already excludes the vulnerable <1.5.0, and the remaining entries are unpinned.

Note these alerts were previously resolved and then re-added. #1968 bumped torch, transformers, nltk, protobuf, requests, gradio and bs4 in this same file but left scikit-learn, py7zr and accelerate untouched — exactly the three packages flagged now.

Changes

Package	Before	After	CVEs
scikit-learn	==1.2.0	>=1.5.0	CVE-2024-5206 (medium) — sensitive data leakage in TfidfVectorizer.stop_words_
py7zr	==0.20.6	>=1.1.3	CVE-2026-23879 (high) arbitrary file write; CVE-2026-55195 (medium) decompression bomb DoS; CVE-2026-55206 (medium) quadratic complexity DoS in PackInfo._read()
accelerate	==0.34.2	>=1.15.0	CVE-2026-69112 (medium)

Lower bounds are used instead of exact pins so future security patches are picked up without another manual bump, consistent with the style adopted for torch/transformers/protobuf in #1968.

Verification

• Resolved the complete requirements set with uv pip compile --python-version 3.10: resolves cleanly with no conflicts, yielding scikit-learn 1.7.2, py7zr 1.1.3, accelerate 1.15.0. Both new floors require Python >= 3.10, which the ToRR flow satisfies.
• No repo code changed. This file is referenced only from prepare/benchmarks/torr/README.md for manual benchmark runs and is not installed by CI.
• Pre-commit hooks pass.

Two caveats for reviewers

CVE-2026-69112 (accelerate) has no upstream fix. The advisory lists no patched version. OSV's "fixed" event points at commit beb0672, which is only the v1.14.0 release bump — it marks the last-affected boundary rather than a fix. Diffing the 1.14.0 and 1.15.0 sdists confirms the vulnerable os.path.join(checkpoint_folder, f) in utils/modeling.py, which joins unvalidated paths taken from the index JSON, is unchanged. So 1.15.0 is the latest release but still technically affected, and Dependabot may continue to flag it. Mitigating factor: unitxt never calls load_checkpoint_and_dispatch directly, reaching accelerate only through transformers' device_map, so the path is not reachable with attacker-controlled checkpoints via unitxt's own API. Bumping moves the pin off the flagged 0.34.2 and will pick up the real fix once released.

pyproject.toml scikit-learn<=1.5.2 left unchanged. It is not vulnerable, and the upper bound appears deliberate — it mirrors the scikit-learn==1.5.2 pin in .github/workflows/test_helm.yml, which runs on Python 3.9 for HELM compatibility. Widening it risks breaking that job for no security benefit, so it seemed better handled separately if desired.

🤖 Generated with Claude Code