SinhalaMMLU-SharedTask/team-rexx-training-data

Team Rexx: training data for the SinhalaMMLU shared task

This repository holds the training data of SinhalaMMLU-SharedTask/team-rexx-gemma-4-e4b, our shared-task model. The training code and the full lineage are in SinhalaMMLU-SharedTask/team-rexx under training/.

The model was trained on 20,870 rows: a Sinhala MCQ, its four options, the keyed answer and an English reasoning trace that ends on that answer. 18,635 rows are machine-written and are published here. The other 2,235 rows are on 930 questions from the official development set, which we do not republish because that dataset is gated.

File	Rows	What it is
train_generated.jsonl	18,635	Every training row whose question was machine-written from curriculum text
train_generated.provenance.json	8,887	The closed model behind each closed-model row in that file
dev_questions_used.jsonl	930	Which official development questions were trained on (file and q_no in naist-nlp/SinhalaMMLU, all easy level and outside the 600 we kept for validation), with the number of traces per question

No question from the shared-task test set is in the training data. A check over all 20,870 rows found 0 test questions with the same question and options.

Fields of train_generated.jsonl

Field	Meaning
id	tg:… = question written by Gemma 4 26B-A4B; gen:… = written by a closed model and blind-checked. The suffix after # names the trace: kd… = open Gemma teacher, cx = Codex, gk = Grok, op = Claude Opus.
subject, grade, level	Curriculum subject, school grade and shared-task level of the source passage
source	kd (trace by an open Gemma teacher, 9,748 rows) or closed (trace by a closed model, 8,887 rows)
question, choices, answer	Sinhala question, four options, keyed answer 1–4
reasoning_en	English reasoning, kept only if it ends on the keyed answer

How it was made

1. Curriculum text. Text was extracted from the official Sinhala-medium textbooks for grades 6–11 (Educational Publications Department) and from the A/L teacher guides, syllabi and resource books (NIE, e-thaksalawa). Legacy FM-font text was converted to Unicode, and garbled passages were filtered out. No past exam papers were used.
2. Questions.
   • 14,571 rows are on questions written by Gemma 4 26B-A4B-it (open weights) from those passages.
   • 4,064 rows are on questions written by OpenAI Codex gpt-6.1-sol (3,968 rows) or by Claude via Claude Code in a pilot round (96 rows).
   • A closed-model question was kept only if a second model, Claude Sonnet or Codex, chose the keyed answer without seeing the key.
3. Reasoning traces.
   • Open traces come from Gemma 4 26B-A4B-it and 31B-it, run offline.
   • Closed traces come from Codex gpt-6.1-sol (8,466 rows here), xAI Grok grok-4.7-build-fast (388) and Claude Code claude-opus-5-5 (33).
   • Every closed trace was answered blind: no key, no passage, tools and web search off.
   • A trace was kept only if all of these held:
     • its answer matched the key;
     • it had 120–900 words;
     • it did not mention a passage, an answer key or a URL.

Closed models were used only to generate this training data, never at inference.

Licence and use

The questions and traces were generated by the models named above from Sri Lankan government curriculum material. They are shared for research on the SinhalaMMLU shared task. Check the terms of the generating models and of the source material before other uses.