Add Indic translations with human-verified corrections

README.md

@@ -0,0 +1,107 @@
+---
+license: mit
+language:
+- as
+- bn
+- gu
+- hi
+- kn
+- ml
+- mr
+- ne
+- or
+- pa
+- sa
+- ta
+- te
+- ur
+size_categories:
+- n<1K
+task_categories:
+- text-generation
+tags:
+- indic
+- multilingual
+- translation
+- benchmark
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: test.parquet
+---
+
+# indic-jee-bench
+
+JEEBench (JEE Advanced 2016-2023) translated into 14 Indian languages, with a human-verified
+correction pass over the translations the automatic judge scored lowest.
+
+Upstream source: [`daman1209arora/jeebench`](https://huggingface.co/datasets/daman1209arora/jeebench)
+
+## Layout
+
+One row per original benchmark item, with every language on that same row. The
+original columns are carried through unchanged; each translated field gains one
+column per language:
+
+```
+question -> question_Hindi_translation, question_Bengali_translation, ...
+```
+
+| | |
+|---|---|
+| rows | 515 |
+| languages | 14 |
+| translated fields | `question` |
+| columns | 21 |
+
+Languages: Assamese, Bengali, Gujarati, Hindi, Kannada, Malayalam, Marathi, Nepali, Odia, Punjabi, Sanskrit, Tamil, Telugu, Urdu.
+
+## How it was built
+
+1. **Translation.** Each field was translated per language by an LLM, then
+   scored by a separate judge model, and anything the judge failed was sent to a
+   correction pass and re-judged.
+2. **Human verification.** 1388 (item, language) instances were
+   reviewed by three annotators each, with a fourth reviewer choosing among the
+   three proposals or writing a final edit. They come from two pools: 148 from
+   the tail the judge failed at least once and sent to correction, and 1240 from
+   a random spot check of instances the judge passed first time. The reviewer's
+   decision is what appears here.
+3. **Structural repair.** 17 reviewer decisions broke something a grader or a
+   renderer depends on (a math delimiter or subscript dropped, a display equation
+   collapsed, an option changed, one question replaced by a different question).
+   In 12 the reviewer's text was kept and only the broken span was restored; in 5
+   the machine translation was restored. A further 29 machine translations had a
+   math delimiter dropped or `\right` corrupted by JSON escaping, and those spans
+   were restored too.
+
+Instances outside the reviewed set carry the machine translation.
+Human review covered 10 of the 14 languages; Assamese, Bengali, Gujarati, Sanskrit were not reviewed and carry the machine translation throughout.
+
+## Caveats
+
+- Answer keys, option counts and any structural field are unchanged from
+  English. For multiple-choice items the option list keeps its original length
+  and order, so the upstream answer key indexes correctly.
+- Scores on this dataset are not directly comparable to the English original:
+  translation quality is part of what is being measured.
+- Only `question` is translated, and it is the whole item: statement, tables and,
+  for multiple-choice items, the options inline. `gold`, `type`, `subject`,
+  `description` and `index` are unchanged.
+- `gold` is a letter string (`B`, `ABD`) or a number, graded against the labels and
+  quantities inside the translated question. Option labels (`(A)`/`[A]`) and
+  paired-list labels (`(I)`-`(IV)`, `(P)`-`(T)`) are kept in ASCII, and the
+  source's hedging ("is/are", "statement(s)") is kept, so `MCQ(multiple)` items
+  do not reveal how many options are correct.
+- JEEBench ships no id column. `id` was added: a deterministic UUID derived from the
+  row's content. `description` + `index` (e.g. "JEE Adv 2019 Paper 1", 37) locate the
+  original paper and question.
+- Upstream LaTeX defects are kept in the English column. One matrix question leaves
+  two options' math unclosed upstream; the translations close them.
+- 2 cells the translation pipeline did not deliver (`question_Odia_translation` and
+  `question_Sanskrit_translation`, one row each) were filled afterwards and are not
+  human-reviewed. They are the pipeline model's own output, rejected only because
+  its JSON left LaTeX backslashes unescaped; the escaping was repaired and both were
+  re-judged at 100/100.
+- License: MIT, inherited from upstream.

test.parquet

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2569d5aa16fd7cdc74bbb9b654697b30be535fbceaacda5af8677776b8b078a8
+size 2737009