Datasets:
Upload README.md with huggingface_hub
Browse files
README.md
ADDED
|
@@ -0,0 +1,32 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
---
|
| 2 |
+
license: cc-by-nc-4.0
|
| 3 |
+
language:
|
| 4 |
+
- si
|
| 5 |
+
task_categories:
|
| 6 |
+
- multiple-choice
|
| 7 |
+
tags:
|
| 8 |
+
- sinhala
|
| 9 |
+
- sinhalammlu
|
| 10 |
+
---
|
| 11 |
+
|
| 12 |
+
# team-qbits MCQ SFT data (v3)
|
| 13 |
+
|
| 14 |
+
Supervised fine-tuning data for our ICTer 2026 SinhalaMMLU Shared Task submission.
|
| 15 |
+
|
| 16 |
+
- `sft_train3.jsonl` — 21,235 prompt/completion examples from 7,151 questions
|
| 17 |
+
- `sft_val3.jsonl` — 721 held-out questions, one example each
|
| 18 |
+
|
| 19 |
+
Derived from publicly available Sri Lankan examination papers (pastpapers.wiki,
|
| 20 |
+
e-thaksalawa.moe.gov.lk), O/L and Grades 6-11, covering 19 subjects including all 14 that
|
| 21 |
+
appear in the SinhalaMMLU easy split. We do not redistribute the source PDFs.
|
| 22 |
+
|
| 23 |
+
**Decontamination.** Checked against the released SinhalaMMLU `TEST/easy` split, which matched
|
| 24 |
+
and removed 26 questions. That split was used only as a blocklist — never for training,
|
| 25 |
+
checkpoint selection, prompt design or hyperparameter choice. The train/validation split is
|
| 26 |
+
grouped by source paper, so no paper appears on both sides.
|
| 27 |
+
|
| 28 |
+
**Format.** Each question is rendered under several prompt templates (Sinhala instruction, bare,
|
| 29 |
+
English instruction) with the option order shuffled; the target is the option number and loss is
|
| 30 |
+
applied to the answer tokens only.
|
| 31 |
+
|
| 32 |
+
Pipeline and code: `SinhalaMMLU-SharedTask/team-qbits`.
|