Skip to content

Sommelier xLAM single-call Hebrew paired rows

Publication template — full Hebrew v3 evidence is pending. Replace this block with a release-specific evidence statement and remove the literal REPLACE_FROM_VERIFIED_DATASET_BUNDLE marker only after the audited rows, semantic gate, and publication manifest exist. The dataset publisher rejects an unresolved marker.

This CC-BY-4.0 dataset is derived from Salesforce/xlam-function-calling-60k. Sommelier filters and selects the root corpus, machine-translates only each natural-language query into Hebrew, and keeps tool schemas and gold answers byte-identical. Every accepted row names its English root with source_example_id and inherits that root's split.

The one-time forward teacher is the OpenAI Responses snapshot gpt-5.5-2026-04-23 using explicit Flex service. The public translation_summary.json records the exact request/runtime identity, returned model and tier, usage, mechanical audit/drop counts, the maximum observed response input-token count, and a calculated public-list-price value. Full publication binds that value to the recorded local ceiling. The ceiling is not an invoice or provider account/project cap. The dated API name is not a provider-weight digest or a guarantee of byte-identical regeneration. store=false is not a Zero Data Retention claim. The source query and bounded selected-tool projection were processed by OpenAI.

The release is a machine-translated survivor corpus: rows that failed the declared structural, placeholder, Hebrew-script, bidi, or protected-span audits are absent and counted. Before publication, a deterministic 200-row sample is back-translated with the pinned Helsinki-NLP/opus-mt-tc-big-he-en checkpoint and reviewed under Sommelier's non-native semantic rubric. The immutable template, finalized decisions, and whole-publication gate are included. This bounded audit does not establish native fluency or full-corpus semantic correctness.

Files

  • rows.he.jsonl: accepted Hebrew paired rows.
  • translation_summary.json: complete translation identity and accounting.
  • translation_publication.json: canonical row/provenance digest binding.
  • translation_semantic_review_template.json: untouched machine-selected review inputs and back-translations.
  • translation_semantic_review.json: finalized review and release decision.

The private raw provider journal is intentionally not distributed. Its SHA-256 and content-free aggregate are bound by the public summary.

Attribution and intended use

Attribute Salesforce for the source xLAM dataset and indicate that Sommelier filtered, split, and machine-translated the queries. Review the upstream dataset card and CC-BY-4.0 terms before redistribution. This release supports the recorded single-tool-call research experiment; it is not a general Hebrew assistant corpus, a native translation benchmark, or production safety data.