abdelstark/sommelier-xlam-single-call-splits-he-hymt-sanitized

Sommelier xLAM single-call Hebrew paired rows (Hy-MT2, sanitized release)

This CC-BY-4.0 dataset is derived from Salesforce/xlam-function-calling-60k. Sommelier filters the source corpus to single-tool-call examples, deterministically splits it, and machine-translates only each natural-language query into Hebrew. The exact training snapshot kept tool schemas and gold answers byte-identical to the English root. For public release, 15 GitHub-PAT-shaped substrings inherited from the synthetic upstream corpus were replaced with [redacted] in the tools field; queries and answers were unchanged. Every accepted row names its English root with source_example_id; downstream preparation deterministically assigns it to that root's train/validation/test split. The published JSONL itself is exposed as one Hugging Face pairs split.

How the translation was produced

The forward translator is tencent/Hy-MT2-1.8B-GGUF (Hunyuan-MT), run locally through ollama from the official Q8_0 GGUF build with greedy decoding (temperature 0, fixed seed). The exact local 1,908,528,192-byte blob has SHA-256 5c3fe0b1408a5ceb0143184ef247b11b579c525f4b02b060e6c851bb76fef1a4. It matches Hy-MT2-1.8B-Q8_0.gguf at immutable repository revision b27182d810fa3ceb6ed04e7c324c54e35c0d209c byte for byte. This run did not directly load the separate Hy-MT2-1.8B-FP8 artifact. Because Hy-MT2 is a pure translator that would otherwise translate argument values, each gold-bearing protected span in the query is masked with a short ASCII sentinel before translation and restored afterward; rows are retried with alternate sentinels and a final unmasked pass.

Every accepted row passes Sommelier's mechanical translation audit:

• every protected span present byte-identically at token boundaries,
• Hebrew as the dominant script (at least half of unprotected letters),
• no unsafe bidirectional control characters, no Unicode replacement or control/format/surrogate code points,
• no reproduction of the translation instruction,
• length within the preparation budget.

translation_summary.json records the model, decoding settings, per-attempt acceptance counts, per-split counts, and the mechanical drop-reason histogram.

Accounting and immutable identity

• Selected roots: 17,000
• Accepted rows: 16,272 (95.72%)
• Accepted by root split: 14,364 train / 961 validation / 947 test
• Drops: 626 missing protected span / 48 malformed spacing / 54 wrong script
• Exact dataset revision used by training (now privately quarantined): c9598f1855ca88e11a4aa31c53ad387faf467eb2
• Exact training rows.he.jsonl SHA-256: 94f1c23022f24065edad136cf4768a21576db4d89a3e45774f7806f69adbcabf
• Public sanitized rows.he.jsonl SHA-256: d7f2895f2b3976cc7ea917dc85633b0140efe03737bda8e20d69ba0c449cacee
• translation_summary.json SHA-256: 0ab7b62615dce11dc705faccf047df105f782665a0e3bbaf157dc46c995bc5ce

The downstream preparation gate removed 86 duplicate translated queries. The training run therefore consumed 14,286 / 955 / 945 Hebrew examples in its train/validation/test splits.

Public security sanitization

The upstream APIGen corpus is gated, research-oriented synthetic data, but 15 rows contained strings matching the classic ghp_ GitHub personal-access-token shape. Sommelier did not test those strings: synthetic-looking credentials are not assumed harmless. The exact training snapshot was therefore quarantined, and the public JSONL replaces all 15 matches with [redacted]. No row was removed; only the tools field changed in those 15 rows. The deterministic public_sanitization.json binds the original revision/hash, pattern classes, replacement counts, and public output hash. Consequently, the public JSONL is not byte-identical to the dataset used by training.

Honesty and limitations

This is a machine-translated survivor corpus. Rows that failed the mechanical audits are absent and counted. The audits are structural: they cannot prove that a translated query preserves the intent of the English root.

There is no human semantic review. This dataset is deliberately distinct from the preregistered Sommelier Hebrew v3 evidence dataset, which uses a different forward teacher and a human-signed back-translation review. This release supports a local, open-model tool-calling experiment; it is not a native Hebrew benchmark, a general Hebrew assistant corpus, or production safety data.

Files

• rows.he.jsonl: accepted Hebrew paired rows (sommelier.raw_tool_call_row.v1: query, tools, answers, source_example_id, source_id, source_revision, schema_version).
• translation_summary.json: translation identity and mechanical accounting.
• translator_identity.json: a post-run immutable binding from the local Ollama blob to the official Tencent GGUF file. It supplements rather than rewrites the translation summary used by training.
• public_sanitization.json: the closed accounting and input/output hashes for the credential-pattern sanitization applied before public release.

Attribution and intended use

Attribute Salesforce for the source xLAM dataset and indicate that Sommelier filtered, split, and machine-translated the queries with Tencent Hy-MT2. Review the gated upstream dataset card, its research/synthetic-data limitations, and CC-BY-4.0 terms before redistribution. Cite Zuxin Liu et al., “APIGen: Automated Pipeline for Generating Verifiable and Diverse Function-Calling Datasets,” arXiv:2406.18518 (2024).