Single val split: 1,013 retained records from 303 source train + 744 source validation records of tridm/VMLU. Source train: 292 retained / 11 excluded; source validation: 721 retained / 23 excluded. All 58 subjects remain represented. No test data was downloaded or included.

Dataset viewer and inputs

Only these nine columns are exported in the Parquet dataset:

Column	Meaning
`id`	Original sample ID. The subject is the ID prefix.
`question`, `choices`, `answer`	Original source question, options and answer, unchanged.
`prompt`	Original instructions, written question and choices for text evaluation, unchanged.
`prompt_audio`	Audio instructions plus the original choices; the question section is removed.
`transcription`	Exact original question, including written notation.
`tts_text`	Complete, natural spoken question; preserves the meaning, data and mathematical scope of the source.
`audio`	Regenerated question-only WAV, mono PCM16 at 16 kHz, embedded in Parquet.

Every prompt_audio begins with:

Nghe bản ghi, sau đó trả lời câu hỏi trắc nghiệm bên duới bằng tiếng Việt.

The remaining answer-letter instructions and all original choices are unchanged. For text evaluation use prompt. For audio evaluation supply prompt_audio together with audio; do not send question, transcription or tts_text as additional model context. The audio contains the question only, without options, answer keys or model instructions.

Spoken language revision

Every retained question was individually reviewed for a complete spoken question. Incomplete written stems are converted to a natural question with the appropriate interrogative expression, such as nơi chốn, thời điểm, số lượng or khái niệm. For example, “Hồ sơ đăng ký hộ kinh doanh được nộp tại:” becomes “Hồ sơ đăng ký hộ kinh doanh được nộp tại đâu?”. This changes tts_text and the regenerated waveform; it does not change the original question, transcription, prompt, options or answer key.

Rewrites preserve numerical values, units, negation, the correct/incorrect/exception selection objective, formula grouping, ordered coordinates, bitstrings and literal quotations. No solution or answer choice is inserted into the spoken question. Existing pronunciation conventions remain: x → ích xì, m → e mờ, n → e nờ; intentional comma pauses may separate phonetic components when necessary. Domain-specific abbreviation and notation normalization is retained. All 1,013 waveforms were regenerated using the revised recipe and checked again after final resampling.

Source membership, subject ID, audio hashes and durations, normalization versions and notes, audio review status and observations are stored separately in docs/sample_metadata.jsonl, keyed by id; they are not dataset-viewer columns. Complete before/after spoken text and semantic review evidence are under docs/spoken_v3/. Current synthesis provenance and audio review are in docs/audio_provenance.jsonl, docs/audio_review.jsonl and docs/audio_v3/. Previous revision evidence remains accessible in the repository history.

Two agents separately reviewed the spoken-language changes and the emitted audio. An independent audio-model transcription receives no source text; a separate comparison checks the expected spoken question against the final waveform. Raw observations, disagreements, repair attempts and final hash-bound decisions are preserved. Signal checks and automated model review do not establish infallible pronunciation. Human listening remains explicitly not_listened. Critical phonetic decisions for 05-0005 and 14-0009 use matching broad and focused audio evidence with preserved conflicting crop transcriptions; they are weighted model decisions, not unanimous or human-certified pronunciation. The 11-0009 final rendering uses shorter segments to retain both spoken semicolons. Its literal int spelling preserves N identity, but the e-prefix of the spoken N name is nonunanimous in cropped ASR; the final check explicitly records moderate-confidence weighted acceptance based on the broad spelling and both full reviews. Its exact synthesis recipe, prior failed waveforms and all conflicting crops remain in the audit files.

Fair text/audio comparison

Text and audio evaluation use the same 1,013 sample IDs and the same 34 source exclusions. Relative to the historical 1,016-case text evaluation, three further IDs are excluded from current text scores as well:

ID	Reason
`05-0003`	Source answer conflicts with the mathematical question and choices.
`21-0017`	The premise says free warranty (bảo hành), while the question says free insurance (bảo hiểm).
`52-0002`	The source wording xử lý khí is inconsistent with the BOD5/COD context and needs source verification.

These are source-quality exclusions shared by both modalities. Provider errors or blocked responses are not extra exclusions and remain in the accuracy denominator. Historical raw model outputs are retained for audit; current text scores are recomputed by filtering the shared excluded IDs, without repeating inference or changing source answer keys.

The Parquet rows, audio, prompts, nine viewer columns, original statistics and retained source counts are unchanged by this documentation update. The full shared policy and eligible-ID fingerprint are in docs/evaluation_policy.json.

Statistics and exclusions

statistics.json and subject_statistics.csv describe the current dataset and losses versus the original. original_statistics.json and original_subject_statistics.csv preserve the prior statistics. download_manifest.json preserves original download hashes. Excluded records and their unchanged questions/choices/answers remain available in docs/excluded_records.jsonl; excluded_samples.csv lists reasons. Source defects requiring external comparison were omitted rather than guessed.

Subject IDs follow the upstream README mapping. The upstream dataset_stat.csv has a different ID mapping.

ID	Subject	Original train	Original validation	Retained train	Retained validation	val	Excluded
01	Elementary Mathematics	6	14	4	13	17	3
02	Elementary Science	6	14	6	14	20	0
03	Middle School Biology	5	13	5	13	18	0
04	Middle School Chemistry	6	14	5	11	16	4
05	Middle School Mathematics	3	8	2	6	8	3
06	Middle School Physics	6	14	6	12	18	2
07	High School Biology	6	14	5	13	18	2
08	High School Chemistry	6	14	6	14	20	0
09	High School Mathematics	4	11	3	9	12	3
10	High School Physics	6	14	6	14	20	0
11	Applied Informatics	6	14	6	13	19	1
12	Computer Architecture	6	14	6	14	20	0
13	Computer Network	5	13	5	13	18	0
14	Discrete Mathematics	5	12	5	12	17	0
15	Electrical Engineering	5	13	5	13	18	0
16	Introduction to Chemistry	5	13	4	11	15	3
17	Introduction to Physics	5	13	5	13	18	0
18	Introduction to Programming	5	13	5	12	17	1
19	Metrology Engineer	4	10	4	10	14	0
20	Operating System	6	14	6	14	20	0
21	Statistics and Probability	5	13	5	12	17	1
22	Middle School Civil Education	5	13	3	13	16	2
23	Middle School Geography	4	11	4	11	15	0
24	High School Civil Education	6	14	6	14	20	0
25	High School Geography	5	12	5	11	16	1
26	Business Administration	5	13	5	12	17	1
27	Ho Chi Minh Ideology	5	13	5	12	17	1
28	Macroeconomics	6	14	6	14	20	0
29	Microeconomics	6	14	6	14	20	0
30	Principles of Marxism and Leninism	6	14	6	14	20	0
31	Sociology	5	13	5	13	18	0
32	Elementary History	5	13	5	13	18	0
33	Middle School History	6	14	6	14	20	0
34	Middle School Literature	5	13	4	13	17	1
35	High School History	6	14	6	13	19	1
36	High School Literature	6	14	6	12	18	2
37	Administrative Law	3	7	3	7	10	0
38	Business Law	5	13	5	13	18	0
39	Civil Law	6	14	6	14	20	0
40	Criminal Law	5	12	5	12	17	0
41	Economic Law	5	12	5	12	17	0
42	Education Law	5	12	5	12	17	0
43	History of World Civilization	6	14	6	14	20	0
44	Idealogical and Moral Cultivation	6	14	6	14	20	0
45	Introduction to Laws	4	9	4	9	13	0
46	Introduction to Vietnam Culture	6	14	6	14	20	0
47	Logic	5	13	5	13	18	0
48	Revolutionary Policy of the Vietnamese Commununist Part	6	14	6	14	20	0
49	Vietnamese Language and Literature	5	13	5	13	18	0
50	Accountant	5	13	5	13	18	0
51	Clinical Pharmacology	6	14	6	14	20	0
52	Environmental Engineering	5	13	4	13	17	1
53	Internal Basic Medicine	5	13	5	13	18	0
54	Preschool Pedagogy	3	7	3	7	10	0
55	Tax Accountant	5	13	5	13	18	0
56	Tax Civil Servant	5	13	5	13	18	0
57	Civil Servant	5	13	5	12	17	1
58	Driving License Certificate	5	13	5	13	18	0

Excluded IDs

ID	Source split	Reason
01-0005	train	Có lựa chọn trùng nội dung sau khi bỏ nhãn và gộp khoảng trắng; giữ nguyên đáp án, không tự sửa.; Chỉ chuyển biểu thức thành lời, không tính kết quả. Hai lựa chọn B/C trùng trong nguồn.
16-0003	train	Mũi tên phản ứng có vẻ bị lỗi thành à; lựa chọn A và B trùng nội dung.; Có lựa chọn trùng nội dung sau khi bỏ nhãn và gộp khoảng trắng; giữ nguyên đáp án, không tự sửa.
09-0001	train	Câu hỏi nhắc bảng biến thiên nhưng row không có bảng/hình.
01-0012	validation	Có lựa chọn trùng nội dung sau khi bỏ nhãn và gộp khoảng trắng; giữ nguyên đáp án, không tự sửa.
11-0014	validation	Cuối câu hỏi chứa A.a=0B.a=-1, nghi lẫn phần lựa chọn vào câu hỏi.
07-0007	validation	Có lựa chọn trùng nội dung sau khi bỏ nhãn và gộp khoảng trắng; giữ nguyên đáp án, không tự sửa.
05-0007	validation	Gọi x_2 là hai nghiệm: nghi thiếu x_1; cần đối chiếu nguồn.
05-0011	validation	Câu hỏi phụ thuộc hình vẽ nhưng row không có hình.
04-0009	validation	Chuỗi \buildrel\emsp\emsp\frac\rightarrow có markup đáng ngờ.
04-0015	validation	Chuỗi \buildrel\emsp\emsp\frac\rightarrow có markup đáng ngờ.
16-0009	validation	B ® A: nghi ký hiệu mũi tên bị lỗi; DH0/DH1 cần xác minh ký hiệu gốc.
16-0016	validation	M[Al(CH ) ] = 72 có vẻ mất chỉ số; cần xác minh nguồn.
09-0005	validation	Có lựa chọn trùng nội dung sau khi bỏ nhãn và gộp khoảng trắng; giữ nguyên đáp án, không tự sửa.
09-0007	validation	Câu hỏi nhắc bảng biến thiên nhưng row không có bảng/hình.
06-0017	validation	Câu hỏi phụ thuộc hình ba ô tô nhưng row không có hình.
06-0020	validation	00 < α < 900: nghi mất ký hiệu độ; cần đối chiếu nguồn, không tự đổi thành 0 đến 90 độ.
04-0013	validation	Corrupted reaction arrow markup and heat-condition notation; no source repair.
01-0004	train	Question depends on a missing bar chart; no visual field exists.
04-0005	train	Question says water contains ions but the ion list is missing.
34-0002	train	Question asks about underlined text, but no underline span is present.
36-0016	validation	Question asks about underlined text, but no underline span is present.
22-0001	train	Source says tộc và nữ instead of a reliable description of the marriage partners; do not silently repair.
22-0002	train	Source says tộc và nữ in a marriage-age question; do not silently repair.
18-0010	validation	Source contains unexplained 7cú pháp; need comparison rather than silently deleting character.
57-0013	validation	Source year 201 1-2020 is broken; do not silently infer 2011.
07-0006	train	Genotype AaXBY has no superscript/span distinguishing X-linked allele from chromosome text; source comparison needed.
26-0016	validation	Job labels include two b labels and no c label; mapping options depend on corrupted labels.
35-0156	validation	Malformed source diacritics in gạ̣p; do not infer corrected vowel/tones for synthesis.
25-0012	validation	Source ngành đóng dầu disagrees with repeated spoken đóng tàu; requires source comparison, no correction guessed.
27-0018	validation	Source từ Anh trả lại Pháp requires comparison; audio recognition corrects to trở lại in context, no repair guessed.
36-0013	validation	Quotation khi lâu versus speech khi đau requires source comparison; no repaired quotation inferred.
52-0002	train	Possible source khí/khi error in wastewater-treatment condition; omitted without guessing a repair.
05-0003	train	Source answer-key conflict: with A={0,1,2}, B={-1,1,4}, option B incorrectly states A\B={0,1}; option D correctly states A∩B={1}, but the source key selects D for the false-statement question. Source fields were not repaired.
21-0017	validation	Source describes free warranty (bảo hành) but asks about free insurance (bảo hiểm); omitted without replacing either term.

Downloads last month: 82

Total file size:

294 MB