arunp0/yama-indic-gemma

Yama Indic typed decisions on real text (Gemma 4 + Qwen3.6)

Typed questions about real Indian-language web text in 11 languages, written by Gemma 4 and verified by Qwen3.6, both open-weights and Apache-2.0. This is the part of Yama v3's training data that is new in v3: 69,465 training questions over 29,673 texts, plus held-out test splits. The texts come from the verified subset of Sangraha (AI4Bharat, CC-BY-4.0). No proprietary-model output was used.

Config	What it is	Train texts	Train questions	Test texts	Test questions	Languages
real_text	topic plus 2–3 snippet-specific questions (choice, noul, score)	18,699	54,198	964	2,846	all 11, extra Odia and Urdu
copa	cause/effect questions in the COPA format: a premise and two alternatives	7,782	7,782	415	415	9 (no Telugu or Gujarati)
boolq	yes/no questions about a passage, about half answered "no"	3,192	7,485	160	381	6 (no Hindi)

train is exactly what Yama v3 was trained on. test holds the rows whose question schema falls in the 5% held-out bucket (MD5 hash of the schema); they were never trained on.

from datasets import load_dataset
copa = load_dataset("arunp0/yama-indic-gemma", "copa")

Companion datasets: arunp0/laya-indic-real-text (the v2 real-text questions, Qwen3) and arunp0/laya-indic-synthetic (synthetic texts, Qwen3).

Fields

Field	Description
text	the text the questions are about: a Sangraha snippet (real_text, boolq) or a premise Gemma wrote from one (copa)
language	Hindi, Bengali, Tamil, Telugu, Marathi, Gujarati, Kannada, Malayalam, Punjabi, Urdu or Odia
questions	{question_id: {"type", "instructions", "criteria"}} in the thx01 schema format, instructions in English
gold	the verified answers: an option name or "NONE_OF_THE_OPTIONS" for choice, a boolean for noul, a 0-based level for score
meta	source, generator (gen), verifier (ver), task, and for copa the Sangraha sentence the premise was written from (src_text)

Pipeline

All generation and verification ran locally with vLLM; no third-party inference provider was used. Gemma was served from the unsloth/gemma-4-26B-A4B-it copy of the same weights.

1. Snippet. Cut from a Sangraha "verified" document. The source corpus was deduplicated line by line against every evaluation set used for Yama (SIB-200/FLORES, MASSIVE, IndicSentiment, IndicCOPA, BoolQ-hi). For this release every premise, alternative and passage was checked again against IndicCOPA and BoolQ-hi by normalised exact match: no overlaps.
2. Generation with google/gemma-4-26B-A4B-it:
   • real_text: a topic from a fixed 16-topic list (options assembled in code; in 7% of cases the correct topic is removed so NONE_OF_THE_OPTIONS is the answer) and 2–3 questions about the snippet.
   • copa: a premise, a "what caused this?" or "what happened next?" question and two alternatives, in the snippet's language.
   • boolq: yes/no questions about a longer snippet, in the passage language.
3. Blind verification with Qwen/Qwen3.6-35B-A3B, a different model family: it answers every question from the text alone, and a question is kept only if it agrees with Gemma. For copa, 58% of generated questions were kept.

Languages were left out on purpose, so that Yama's benchmark gains there measure transfer rather than format matching: copa has no Telugu or Gujarati, and boolq has no Hindi.

Limitations

• Every label comes from two language models agreeing, not from humans. Agreement filtering removes many errors, but not the ones both models share.
• Sangraha is mostly news and web prose, not chat or social media.
• boolq covers only 6 languages: verification of the Malayalam, Punjabi, Urdu and Odia rows did not finish in time.

Licence

CC-BY-4.0, following Sangraha. Credit AI4Bharat for the source texts. Gemma 4 and Qwen3.6 are Apache-2.0.

Copyright 2026 Arun Prabhakaran.