spandyie/amadablam-dpo-preferences

Ama Dablam DPO Preference Data

Preference pairs used to DPO-tune Ama Dablam, a 322M trilingual (Nepali/Maithili/Bhojpuri) language model, across all three languages and three writing systems (Devanagari, IAST, phonetic romanization). See the technical report §9 for full methodology.

Splits

Fields

Each row: {"lang": "ne"|"mai"|"bho", "script": "deva"|"iast"|"phon", "source": ..., "kind": "translated"|"length"|"script", "prompt": str, "chosen": str, "rejected": str}

• —lang / script — language and script the whole triple (prompt+chosen+rejected) is rendered in (Devanagari 85% / IAST 5% / phonetic 10%, content-hash assigned).
• —kind — translated (machine-translated English preference pairs from HuggingFaceH4/ultrafeedback_binarized and argilla/distilabel-intel-orca-dpo-pairs, IndicTrans2 for ne/mai, NLLB for bho, chrF≥60 back-translation gate), length (programmatically degraded run-on/truncated negatives), or script (script-fidelity negatives — a correct response rendered in the wrong script).
• —source — finer-grained provenance tag within kind (e.g. ultrafeedback, argilla_orca, length_runon, script_synth_iast).

Text fields are plain UTF-8 strings — not tied to Ama Dablam's tokenizer. This is a decoded, human-readable mirror of the exact tokenized shards the model trained on (built by decoding the training token IDs back through the source tokenizer, not regenerated separately), so it matches training data exactly.

Provenance and known limitations

• —Translated-pair targets were 10k/language originally; the real measured MT keep rate (~20% of attempted candidates passing the chrF/length/dedup/placeholder-leak gates) made that an ~8-day continuous-GPU undertaking, so targets were revised down: translatedne 2,484 / translatedmai 2,498 / translated_bho 3,940.
• —Maithili and Bhojpuri have no native instruction-tuning source; their length/script pairs are pivot-translated (ne→en→target) from the same Nepali SFT prompt pool used elsewhere in this project, reusing the same MT backends as the translated pairs (no new model checkpoints).
• —Two rows found post-build with un-transliterated glued English inside otherwise-Maithili text (a genuine MT failure not caught by the existing placeholder-leak/code-detection filters) were removed before this release; the rest of all three languages' translated files were audited for the same pattern and no further instances were found.
• —Bhojpuri's kind=translated pairs route through NLLB (no IndicTrans2 support for Bhojpuri), which is CC-BY-NC-4.0 and is why this dataset is released under the same license, non-commercial.
• —IAST and Bhojpuri are the thinnest/lowest-quality legs of this dataset (see the technical report's §9.5 discussion of the resulting preference-accuracy gap) — treat any downstream model trained on this data as inheriting that imbalance.

License

CC-BY-NC-4.0 (non-commercial) — inherited from the Bhojpuri NLLB translation backend and, for the pivoted length/script pairs, from the same Alpaca/Dolly-family lineage as Ama Dablam's SFT data.