👁 Image

madiedgar

Legesher org May 25

Brings the LoRA repo README into alignment with the canonical project source-of-truth at legesher/language-decoded-experiments (HF PR #43, merged 2026-05-25).

Key changes

• Adapter inventory rewritten to reflect actual subdirectory structure: cond-1-en (5k @ 3 seeds, 20k @ 1 seed), cond-2-{zh,es,ur} (5k @ 3 seeds, 20k @ 1 seed), cond-3-zh-5k-native-code (1 seed), cond-5-{zh,es,ur}-5k-c4ai-aya-expanse-32b (1 seed each). Adds cond-5 — the previous README had no cond-5 entries.
• Cond-2 reframed as "Reserved-Word Translation (Legesher)" — Legesher v0.7.3 translates Python's reserved words (keywords, exceptions, built-in functions, numerical system for some target languages) into the target language; user logic preserved.
• Cond-5 sequencing clarified: Legesher transpilation runs first (reserved words), then c4ai-aya-expanse-32b via the Cohere API translates the remaining content (identifiers, comments, docstrings, string literals).
• Cond-4 reframed as "Community-Contributed Native Code" (pending sufficient contributions; no adapter exists yet).
• Source-file control callout added: cond-1, cond-2, cond-5 share the same 5k file subset from bigcode/the-stack-v2-dedup. Cond-3 is the deliberate exception.
• Why Tiny Aya as the base model framing added.
• Refined-extractor banner added at the top (matches the banner on the experiments dataset).
• Benchmark table updated: SIB-200 and Belebele added; MGSM dropped (with the null-result rationale: 3.35B params + 250 examples + ~5% accuracy floor + ±2.7pp CI = condition differences within noise).
• Experimental ladder narrative rewritten with the new framings (baseline→1, 1→2, 2→3, 2→5, 3→5).
• Citation title updated to "Language Decoded: Exploring the Impact of Native Code on Multilingual Models".
• Limitations refreshed: per-language fine-tuning noted (combined-language training is a planned future condition); 103k variants exist on the data repo but no 103k adapters trained yet; extractor coverage bullet pointing at the refined-extractor work.
• Title callout added at the top with the proposed paper title.

No adapter weights or training configs touched — README-only change. YAML frontmatter preserved verbatim.