Update readme

README.md

@@ -13,20 +13,29 @@ tags:
 
 Supervised fine-tuning data for our ICTer 2026 SinhalaMMLU Shared Task submission.
 
-- `sft_train3.jsonl` — 21,235 prompt/completion examples from 7,151 questions
-- `sft_val3.jsonl` — 721 held-out questions, one example each
+* `sft_train3.jsonl` — 21,235 prompt/completion examples from 7,151 questions
+* `sft_val3.jsonl` — 721 held-out questions, one example each
 
 Derived from publicly available Sri Lankan examination papers (pastpapers.wiki,
-e-thaksalawa.moe.gov.lk), O/L and Grades 6-11, covering 19 subjects including all 14 that
-appear in the SinhalaMMLU easy split. We do not redistribute the source PDFs.
+e-thaksalawa.moe.gov.lk), O/L and Grades 6-11, covering 19 subjects including all 14
+that appear in the SinhalaMMLU easy split. We do not redistribute the source PDFs.
 
-**Decontamination.** Checked against the released SinhalaMMLU `TEST/easy` split, which matched
-and removed 26 questions. That split was used only as a blocklist — never for training,
-checkpoint selection, prompt design or hyperparameter choice. The train/validation split is
-grouped by source paper, so no paper appears on both sides.
+**Extraction.** Grade 6-11 papers are only available as scanned images, so Gemini 3.1 Pro
+was used offline to transcribe those page images into structured MCQ fields. This is OCR
+and layout parsing, not content generation: no question, option or answer key was
+generated, translated, paraphrased or rewritten by any model. Every item is a verbatim
+transcription of a published examination question with its official answer key.
 
-**Format.** Each question is rendered under several prompt templates (Sinhala instruction, bare,
-English instruction) with the option order shuffled; the target is the option number and loss is
-applied to the answer tokens only.
+**Decontamination.** Checked against the released SinhalaMMLU `TEST/easy` split, which
+matched and removed 26 questions. That split was used as a blocklist and, separately, to
+choose between already-trained configurations — never for training itself: no gradient was
+ever computed on it, and no question, option or answer key from it entered the training
+set, the CPT corpus or the prompts. We used no part of the medium or hard splits for any
+purpose. The train/validation split is grouped by source paper, so no paper appears on
+both sides.
+
+**Format.** Each question is rendered under several prompt templates (Sinhala instruction,
+bare, English instruction) with the option order shuffled; the target is the option number
+and loss is applied to the answer tokens only.
 
 Pipeline and code: `SinhalaMMLU-SharedTask/team-qbits`.