Datasets:

---

legesher

/

language-decoded-data

like 0

Follow

Legesher 6

Brings the training-data dataset card into alignment with the canonical project source-of-truth at legesher/language-decoded-experiments (HF PR #43, merged 2026-05-25). Targeted edits — the YAML frontmatter (lines 1-906) and the schema tables are unchanged.

Key changes

• Title callout added at the top with the proposed paper title ("Language Decoded: Exploring the Impact of Native Code on Multilingual Models").
• Research question rewritten with the new differential-impact framing (how does the impact of non-English code differ from English code).
• Dataset description updated to use the new condition naming (cond-1 raw English, cond-2 Legesher-transpiled reserved-word translation, cond-3 community-collected mixed native sources, cond-4 community-contributed native authoring, cond-5 Legesher + c4ai-aya-expanse-32b full file translation).
• Source-file control callout added (NEW subsection): cond-1, cond-2, and cond-5 share the same 5k/20k file subset from bigcode/the-stack-v2-dedup. Cond-3 is the deliberate exception with a different source population (and may include non-Python files by design).
• Cond-5 sequencing made explicit: Legesher transpilation first (reserved words), then c4ai-aya-expanse-32b via the Cohere API translates the remaining content (identifiers, comments, docstrings, string literals). Logic and structure preserved throughout.
• Available Configs table terminology cleaned up: "Keyword-swapped Python via Legesher v0.7.3" → "Legesher v0.7.3 reserved-word translation"; cond-3 row changed from "Blended: native Chinese code + transpiled Python" to "Community-collected raw native sources"; cond-4 row changed from "Strictly native Chinese code" to "Native-authored Chinese code (community-contributed)"; cond-5 rows changed from "Model-translated (full LLM translation via Cohere Aya)" to "Legesher transpilation + full c4ai-aya-expanse-32b translation".
• Cond-3 schema description updated to reflect the Phase 2 → Phase 3 meaning shift of "Mixed Native Sources": Phase 2 sense was a planned cond-2 + native composite; Phase 3 sense is varied online source locations from real-world repositories.
• Cond-4 schema description updated with the "Community-Contributed Native Code" framing — design goal is code whose problem-solving logic is itself native; cond-5's translated data served as the Phase 3 practical proxy.
• Experimental Conditions table updated with new condition names and descriptions across the board.
• Experimental Ladder rewritten with the new comparison framings (baseline→1, 1→2, 2→3, 2→5, 3→5).
• Limitations terminology cleaned up: "Keyword-only transpilation" → "Reserved-word-only transpilation (cond-2)"; Python-only constraint clarified per condition (cond-1/2/5 Python-only; cond-3 may include non-Python by design); cond-4 framing aligned with the "logic in the target language" goal.
• Citation title updated.

No JSON, parquet, schema rows, or YAML frontmatter touched — README prose body only. The dataset_info splits, configs, and column schemas are unchanged.