Arushhh/indic-translate-mar-bhili-lora

Marathi ↔ Dehwali Bhili LoRA

LoRA adapter over `bodhan-ai/indic-translate`, a derivative work used under the Indic Open Model License v1.0.

Bhili is not one of the base model's 22 supported languages, so this teaches the model a language it has never seen rather than adapting one it knows. Note that extra_languages extends the toolkit's renderer but not its inference engine, which raises ValueError: unsupported language 'Bhili' — the engine's language table has to be extended separately for the served prompt to match the trained one.

Data

AI Kosh Dehwali Bhili Krishi Darshini, 21,622 validated pairs (agricultural extension material), plus numeral-substitution augmentation and conversational rows from misniitdelhi/AdiBhasha.

Measured properties of the corpus that shaped the recipe

• —Reference numeral-preservation ceiling: 83.7%. 9.5% of references verbalise numerals (20 → वीस, 800 → आठहोव) rather than dropping them. This is a Bhili prose convention, not corruption — but the prompt contract is frozen, so there is no channel to request one style over the other.
• —Copy prior: 33% exact-copy at 1–2 source words, 1.2% at 21+. Bhili is lexically close to Marathi, so "echo short inputs" is a cheap rule that fits the training distribution and fails on exactly the short phrases that matter.
• —Zero greetings in 21,622 rows (धन्यवाद, नमस्कार, स्वागत all absent).

Evaluation

Reported against a COPY baseline, because copying preserves numerals perfectly by construction and scores respectably on chrF++ given the languages' closeness. Metrics: chrF++ (word_order=2), numeral-sequence exact match, digit-script match, exact match, and spurious copy rate — the fraction of outputs that merely echo the source where the reference differs from it.