hungphongtrn/udm-eval-v1

UDM Eval v1

Frozen evaluation suite ("v1") for UDM decision models: models that read a state and return a probability for every option of every question asked about it. Vietnamese first, with some English tracks.

• Test splits only. Never train on these rows. Every training source is decontaminated against them (see Provenance and protection).
• Release revision: 5ef90d8e7f1595ecb0ffcbb44ba40d681cbf89e5.
• Built 2026-10-20 (built_utc in manifest.json: 09:38 to 17:50 UTC, one entry per family), sampling seed 0.
• 13 families, 44 track files, 58,366 records (rows). One of the 44 files, model_routing/routerbench, is kept for protection and reference and is not part of the 43 scored tracks.
• Private dataset. It mixes sources with different licences, several of them unknown (see Licence).

The code that builds and scores it is the udm repository (udm/data/build.py, udm/infer.py, udm/score.py); this card describes what the code does, and manifest.json / stats.json in this repo record the exact pins.

Files

manifest.json                      sources (hf_id, revision, licence, origin, note), seed, per-file rows/groups/sha256, track list
stats.json                         per-track label balance, language, origin, option counts, strata
test/<family>/<track>.jsonl.gz     one gzip JSONL file per track

Each track is one config, named after the track (44 configs). There is no default config, because tracks have different fields, so name the track:

from datasets import load_dataset

ds = load_dataset("hungphongtrn/udm-eval-v1", "aiv-sa24-easy", split="test")
ds[0]["questions"]["label"]["criteria"]   # {'A': 'tiêu cực', 'B': 'trung tính', 'C': 'tích cực'}

Loaded rows equal the raw JSON lines (checked for every track with datasets 5.1.0), with one exception: in deepseek-v4-pro, metadata.score is rounded to 10 decimal places in 25 of 29 rows. That field is not used for scoring. To get the bytes the manifest hashes, read the .jsonl.gz file directly (hf_hub_download then gzip.open). The dataset viewer is off, because the families do not share one table schema.

To load all 43 scored tracks:

from datasets import get_dataset_config_names, load_dataset

tracks = [c for c in get_dataset_config_names("hungphongtrn/udm-eval-v1") if c != "routerbench"]
suite = {t: load_dataset("hungphongtrn/udm-eval-v1", t, split="test") for t in tracks}

Record format

One line per record: a state and one or more questions about it. Example, test/sentiment/aiv-sa24-easy.jsonl.gz (first row):

{
  "id": "aiv:SA24_easy:0b89fe71...ae4de04",
  "family": "sentiment",
  "split": "test",
  "state": {"text": "bài giảng hay và dễ hiểu ."},
  "questions": {
    "label": {
      "type": "choice",
      "instructions": "Classify the sentiment or emotion expressed by the text using the listed Vietnamese labels.",
      "criteria": {"A": "tiêu cực", "B": "trung tính", "C": "tích cực"}
    }
  },
  "expected": {"label": "C"},
  "metadata": {
    "source": "aiv", "origin": "native", "lang": "vi",
    "group_id": "aiv:SA24_easy:0b89fe71...ae4de04",
    "stratum": "tích cực", "aiv_task": "SA24_easy"
  },
  "_evaluation": {
    "run_id": "aiv-sa24-easy:aiv:SA24_easy:0b89fe71...ae4de04",
    "dataset": "aiv", "track": "aiv-sa24-easy",
    "group_id": "aiv:SA24_easy:0b89fe71...ae4de04",
    "payload_sha256": "5db9b28647d1522ad41e3bfed7e8099ef3ed663a21718ed0fa2ccc563b75ad72"
  }
}

Field	Meaning
id	Unique record id.
family, split	Task family and test.
state	The input every question is asked about. Keys depend on the track: text, question, prompt, claim + evidence_text, query, tools + messages, conversation, student_submission, and so on. Values are text; tools and messages are stringified lists.
questions	Map of question key to a question. A record can hold several questions about one state.
questions.<key>.type	choice (pick one option) or noul (a yes/no question, no options; criteria is {}).
questions.<key>.instructions	The prompt for that question.
questions.<key>.criteria	The options. Keys are the answer codes (A, B, ...) or, for some tracks, named options (see below). Values are the option text.
expected	The gold answer per question key: an option key for choice, a boolean for noul.
metadata	source, origin (native, translated or synthetic), lang (vi, en; multi for RouterBench), group_id (the unit that sampling and scoring keep together), stratum (the sampling stratum), plus source-specific fields.
scoring	Only on some tracks (reranking, routing, ordinal tracks); see below.
_evaluation	run_id (the key a result must carry: <track>:<id>), dataset, track, group_id, and payload_sha256, a hash of the payload an engine receives (the Decision Index convention).

Question types by family

Family	Questions in a record	How expected looks
sentiment, nli, fact_checking, intent, multiple_choice, education (MC)	One choice question (label, verdict, intent, answer) with a few to 60 options (number varies by track) coded A, B, ...	{"answer": "C"}
education (grading)	grading: noul for deepedubench-grading-binary; for -grading-score an 11-way choice over the scores 0 to 10 (option text is the score, scoring.ordinal is set)	true/false, or a code such as "K" (score 10)
quality_scoring	quality: choice with options keyed "0" to "4" (educational value), scoring.ordinal set	{"quality": "0"}
content_moderation	toxic and contains_hate_or_offense: noul; ViHSD label: 3-way choice	{"toxic": true}, {"label": "CLEAN"}
safety_guardrail	harmful: noul. Some tracks add questions on the same prompt: deepedu attack_type (5-way choice), visafe category (9-way choice), seasafeguard response_harmful (prompt + response), respan one noul per category (jailbreak, system_prompt_leak, unsafe, hate, self_harm, sexual, violence, ...)	{"harmful": true, "category": "cybercrime"}
tool_selection	next_tool: choice over the available tools plus none. Option keys are the tool names	{"next_tool": "find_longest_word"}
skill_routing	skill: choice over skill ids. Option keys are skill ids	{"skill": "skill:monitoring-observability"}
model_routing	route: choice over 11 model names (RouterBench). scoring.type = "routing_outcome" holds per-model quality and cost_usd and the accepted set	{"route": "mistralai/mistral-7b-chat"}
passage_reranking	One noul question per candidate passage, keyed doc:<id>, asking whether the passage helps answer the query. A pool is split into chunks of at most 25 candidates that share metadata.group_id. scoring.type = "retrieval_ranking" holds qrels (graded judgments), field_to_document, scorable_ids, relevance_threshold, fully_judged, combine_chunks_by	{"doc:8182159": true, ...}

Rows of a passage_reranking track are chunks: 17,904 rows come from far fewer queries (the groups column below).

How scoring works

A model gets each record's state and questions and must return, for every question, a probability over its options: a distribution over the keys of criteria for choice, and {"true": p, "false": 1-p} for noul. The reference runner udm.infer writes one line per record, {"run_id", "id", "track", "family", "probs": {question: {option: p}}, "skipped": [...]}. udm.score reads these and does the rest. Models only need to produce probs; the prediction rule is fixed by the scorer:

• choice: prediction is the argmax option.
• noul: prediction is true when p(true) >= 0.5.
• A question missing from the results (not run, or listed in skipped because it was oversized) counts as wrong, as not kept and unranked, and is counted under missing.

Metrics per track

For a track, udm.score computes metrics for each question key and takes the headline from the first question that has a gold answer (for respan that is jailbreak; for visafe, seasafeguard and deepedu it is harmful). Other questions are reported but do not enter the headline.

Question kind	Reported	Headline
choice	accuracy, macro-F1 over the gold labels present, mean NLL; MAE in levels when scoring.ordinal	accuracy
noul	accuracy at p(true) >= 0.5, macro-F1 over {true, false}, mean NLL	macro-F1
retrieval_ranking	nDCG@10 (graded qrels, ideal ranking over every judged passage, so gold missing from the pool counts against), MRR@10; on fully judged pools also keep-set precision, recall and F1 at p(true) >= 0.5	nDCG@10
routing_outcome	accepted-route accuracy, mean quality, mean cost, oracle quality	accepted-route accuracy

Reranking: chunks of one query (combine_chunks_by = metadata.group_id) are merged into one pool, passages are ranked by p(true), and nDCG@10 and MRR@10 are averaged over queries that have at least one relevant judgment.

There is no chance correction. Accuracy and macro-F1 are raw, so the floor differs by track (about 0.5 for balanced binary tracks, 0.25 for 4-option multiple choice, near 0 for 60-way intent, and the majority-class rate on skewed tracks such as victsd with 11% toxic). Compare a model against other models on the same track, not against zero.

Aggregation

Implemented in udm/score.py (summary) and scripts/r2_history.py:

1. Each track has one headline in [0, 1] (table below).
2. udm.score writes metrics.json with every track and with unweighted means of track headlines per family and per pillar, each split by language (<pillar>/<lang>, with the language taken from metadata.lang). It scores every */*.jsonl.gz in the suite directory, so with all 44 files it includes RouterBench.
3. scripts/r2_history.py (the R2 reporting script) selects the 43 active tracks explicitly (all except model_routing/routerbench) and defines the v1 score as the plain mean of those 43 track headlines (the "v1 equal-43-track mean"). Each track counts once, whatever its size, so a 24-row track weighs the same as a 6,680-row track. It also reports the pillar-language group means and a family macro mean. No weights, no rescaling.
4. Pillar-language groups (8 groups, 43 tracks):

Pillar / language	Tracks	Count
agentic_routing/en	skillgym-sft, nemotron-tools, toucan-kimi-k2, deepseek-v4-pro	4
retrieval_grounding/en	trec-dl19, trec-dl20, wixqa-expertwritten, wixqa-simulated	4
guardrail_moderation/en	budguardrail, respan	2
retrieval_grounding/vi	8 Vietnamese reranking tracks (zalo-legal, viequad-retrieval, vinli-reranking, table-markdown-vn, tvpl-retrieval-vn, askubuntu-reranking-vn, stackoverflow-reranking-vn, scidocs-reranking-vn) plus aiv-nli500, aimy144-fact-checking-vi	10
language_understanding/vi	7 multiple_choice tracks, 2 sentiment tracks, massive-vi-vn	10
guardrail_moderation/vi	vihsd, vihos, victsd, deepedu, visafe, aiv-jbb, seasafeguard, multijail	8
preference_quality/vi	fineweb-c-vi	1
education/vi	deepedubench-qa, -cdpk, -grading-binary, -grading-score	4

Family-to-pillar mapping (udm/data/tracks.py): tool_selection, skill_routing, model_routing -> agentic_routing; passage_reranking, nli, fact_checking -> retrieval_grounding; multiple_choice, sentiment, intent -> language_understanding; safety_guardrail, content_moderation -> guardrail_moderation; quality_scoring -> preference_quality; education -> education. The pillars are documented in docs/benchmark_pillars.md of the repo. Per that document, results are reported per pillar with group-bootstrap intervals, and small tracks count through their pillar. The 43-track mean is a summary, not a replacement for the pillar breakdown.

Tracks

Headline: accuracy for choice questions, macro-F1 for noul questions, nDCG@10 for reranking, accepted-route accuracy for routing. "Rows" is the number of records in the file; for reranking and tool/skill tracks, groups (queries, conversations) are given when they differ. Origin is from stats.json. Source revisions are shortened to 8 characters; full values are in manifest.json. routerbench is the 44th file and is not one of the 43 scored tracks.

Family	Track	Lang	Origin	Rows	Source @ revision	Licence	Headline
content_moderation	vihsd	vi	native	6,680	visolex/ViHSD @ eb9ed6e0	unknown	accuracy
content_moderation	vihos	vi	native	1,106	phusroyal/ViHOS @ d465cb3b	mit	macro-F1
content_moderation	victsd	vi	native	1,000	tarudesu/ViCTSD @ 65a073f2	unknown	macro-F1
education	deepedubench-qa	vi	native	795	private AIV / DeepEdu snapshot	unknown (shared by the owner)	accuracy
education	deepedubench-cdpk	vi	translated	900	private AIV / DeepEdu snapshot	unknown (shared by the owner)	accuracy
education	deepedubench-grading-binary	vi	native	165	private AIV / DeepEdu snapshot	unknown (shared by the owner)	macro-F1
education	deepedubench-grading-score	vi	native	63	private AIV / DeepEdu snapshot	unknown (shared by the owner)	accuracy
fact_checking	aimy144-fact-checking-vi	vi	native	7,521	aiMy144/vietnamese-fact-checking-verifier-data-v3-1 @ 6bba4242	not specified in the dataset metadata	accuracy
intent	massive-vi-vn	vi	translated	2,974	AmazonScience/massive @ ed58ac42	CC-BY-4.0	accuracy
model_routing	routerbench	multi	native	1,001 (1,000 groups)	withmartian/routerbench @ 78402148	unspecified by dataset repository	accepted-route accuracy
multiple_choice	aiv-vlmu-mcqa-extend400	vi	native	514	private AIV / DeepEdu snapshot @ 3949bf7e	not specified by the local harness snapshot	accuracy
multiple_choice	aiv-hle-mc-vi	vi	translated	506	private AIV / DeepEdu snapshot @ 3949bf7e	not specified by the local harness snapshot	accuracy
multiple_choice	aiv-culture100	vi	native	100	private AIV / DeepEdu snapshot @ 3949bf7e	not specified by the local harness snapshot	accuracy
multiple_choice	aiv-pragmatic50	vi	native	50	private AIV / DeepEdu snapshot @ 3949bf7e	not specified by the local harness snapshot	accuracy
multiple_choice	aiv-syntax50	vi	native	50	private AIV / DeepEdu snapshot @ 3949bf7e	not specified by the local harness snapshot	accuracy
multiple_choice	mmlu-prox-vi	vi	translated	3,000	li-lab/MMLU-ProX @ 8e6106a6	MIT	accuracy
multiple_choice	vimmrc-vi	vi	native	512	ura-hcmut/ViMMRC @ fe68800e	CC-BY-NC-ND-4.0	accuracy
nli	aiv-nli500	vi	native	500	private AIV / DeepEdu snapshot @ 3949bf7e	not specified by the local harness snapshot	accuracy
passage_reranking	zalo-legal	vi	native	788	GreenNode/zalo-ai-legal-text-retrieval-vn @ 12d76d4d	MIT	nDCG@10
passage_reranking	trec-dl19	en	native	109 (43 groups)	whybe-choi/trec-dl-2019 @ 245ccf2c	unspecified	nDCG@10
passage_reranking	trec-dl20	en	native	142 (54 groups)	whybe-choi/trec-dl-2019 @ 245ccf2c	unspecified	nDCG@10
passage_reranking	wixqa-expertwritten	en	native	800 (200 groups)	Wix/WixQA @ d662dc42	MIT	nDCG@10
passage_reranking	wixqa-simulated	en	native	800 (200 groups)	Wix/WixQA @ d662dc42	MIT	nDCG@10
passage_reranking	viequad-retrieval	vi	native	4,000 (1,000 groups)	mteb/VieQuADRetrieval @ f956535e	MIT	nDCG@10
passage_reranking	vinli-reranking	vi	native	363	ContextSearchLM/ViNLI_reranking @ 84a69290	unspecified	nDCG@10
passage_reranking	table-markdown-vn	vi	native	4,000 (1,000 groups)	GreenNode/GreenNode-Table-Markdown-Retrieval-VN @ 699949f7	MIT	nDCG@10
passage_reranking	tvpl-retrieval-vn	vi	native	4,000 (1,000 groups)	GreenNode/TVPL-Retrieval-VN @ 67c901b6	CC-BY-SA-4.0	nDCG@10
passage_reranking	askubuntu-reranking-vn	vi	translated	305	GreenNode/askubuntudupquestions-reranking-vn @ 5cfaa5c0	CC-BY-NC-SA-4.0	nDCG@10
passage_reranking	stackoverflow-reranking-vn	vi	translated	1,497 (1,000 groups)	GreenNode/stackoverflowdupquestions-reranking-vn @ 3ceb17db	CC-BY-NC-SA-4.0	nDCG@10
passage_reranking	scidocs-reranking-vn	vi	translated	1,100 (1,000 groups)	GreenNode/scidocs-reranking-vn @ c9ab36ae	CC-BY-NC-SA-4.0	nDCG@10
quality_scoring	fineweb-c-vi	vi	native	945	data-is-better-together/fineweb-c @ 913fc625	ODC-By v1.0 (CommonCrawl Terms of Use)	accuracy
safety_guardrail	deepedu	vi	native	961	private AIV / DeepEdu snapshot @ 8e579402	unknown	macro-F1
safety_guardrail	visafe	vi	native/synthetic/translated	1,000	nvidia/Nemotron-Content-VISafe-v1 @ c0d65ade	nvidia-sample-data-license	macro-F1
safety_guardrail	budguardrail	en	synthetic	2,000	budecosystem/guardrail-training-data @ 4924c6b2	mixed-source-licenses	macro-F1
safety_guardrail	respan	en	native	2,000	respanai/guardrail-benchmark @ ee8f386f	mixed-source-licenses	macro-F1
safety_guardrail	aiv-jbb	vi	translated	180	private AIV / DeepEdu snapshot @ 3949bf7e	unknown	macro-F1
safety_guardrail	seasafeguard	vi	native/translated	1,231	suwaimyo/seasafeguard-general-prompt-vie-classification @ ee81f25a	unknown (follows aisingapore/SEASafeguardBench, whose card states none)	macro-F1
safety_guardrail	multijail	vi	translated	315	DAMO-NLP-SG/MultiJail @ 93d65c77	mit	macro-F1
sentiment	aiv-sa24-easy	vi	native	24	private AIV / DeepEdu snapshot @ 3949bf7e	not specified by the local harness snapshot	accuracy
sentiment	aiv-sa26-hard	vi	native	26	private AIV / DeepEdu snapshot @ 3949bf7e	not specified by the local harness snapshot	accuracy
skill_routing	skillgym-sft	en	synthetic	1,527 (1,500 groups)	reasonwang/skillgym-sft @ 51796b33	not specified by the dataset	accuracy
tool_selection	nemotron-tools	en	synthetic	2,585 (1,000 groups)	nvidia/Nemotron-SFT-Agentic-v2 @ 7c804833	CC-BY-4.0; Apache-2.0; MIT	accuracy
tool_selection	toucan-kimi-k2	en	synthetic	202 (45 groups)	Agent-Ark/Toucan-1.5M @ 0df3cf37	Apache-2.0	accuracy
tool_selection	deepseek-v4-pro	en	synthetic	29 (4 groups)	zake7749/deepseek-v4-pro-agent-tool-calling-trajectory @ d6821ad0	not specified by the dataset	accuracy

Notes:

• The private AIV tracks come from aiforvietnam/aiv-benchmark-harness (private, pinned at 3949bf7); DeepEduBench comes from a file shared by the owner. Neither licence is recorded.
• deepedu (safety guardrail) is a separate DeepEdu safety set, not DeepEduBench education.
• trec-dl19 and trec-dl20 both read queries and graded qrels from the pinned mirror whybe-choi/trec-dl-2019; their candidate passages come from the TREC DL BM25 runs.
• The seasafeguard track is read from the ungated mirror suwaimyo/seasafeguard-general-prompt-vie-classification.
• visafe mixes native, synthetic and translated rows; seasafeguard mixes native and translated.
• aiv-hle-mc-vi and mmlu-prox-vi are translations of Decision Index benchmarks (HLE, MMLU-Pro); treat them as a Vietnamese view of those, not as independent evidence (docs/benchmark_pillars.md).

How to run it

Run in the udm repository on a machine with a GPU. Download this dataset so that the test/ directory is local, for example to data/v1/:

python -m udm.infer --adapter d3 --model vllm-sr/d3-mini --suite data/v1/test --out runs/eval/d3-mini
# several GPUs: CUDA_VISIBLE_DEVICES=1 python -m udm.infer ... --shard 1 --shards 3
python -m udm.score --suite data/v1/test --results runs/eval/d3-mini      # writes runs/eval/d3-mini/metrics.json

--families and --tracks restrict a run. udm.infer resumes an interrupted run. udm.score prints one line per track (pillar, family, track, language, headline) and writes per-track, per-family and per-pillar numbers. For the 43-track mean over a set of models, use python scripts/r2_history.py MANIFEST.json --out runs/r2/history-report, which also checks that result run ids match the frozen rows.

To rebuild the files from the pinned sources (needs access to the private AIV and DeepEdu data and to gated sources): python -m udm.data.build --out data/v1 --roles test.

Provenance and protection

• Sources. Each track is one pinned split of one source (hf_id, revision, licence, origin, lang, url and a note on how the split was selected are in manifest.json; viequad-retrieval uses validation). One source split can feed several tracks (TREC DL 2019/2020, WixQA, AIV tasks).
• Sampling. Large tracks are sampled once, by whole groups (metadata.group_id, so all chunks of a reranking query stay together), stratified by metadata.stratum (a per-stratum floor of 50 groups, reduced when the cap cannot afford it), seeded with seed = 0. Capped tracks (cap in the manifest): visafe 1,000; budguardrail 2,000; respan 2,000; nemotron-tools, toucan-kimi-k2 1,000; skillgym-sft 1,500; routerbench 1,000; zalo-legal, viequad-retrieval, table-markdown-vn, tvpl-retrieval-vn, stackoverflow-reranking-vn, scidocs-reranking-vn 1,000 queries; mmlu-prox-vi 3,000. All other tracks keep the full source split. manifest.json gives source_rows and groups per file; stats.json gives the label balance, option counts and strata per track.
• Integrity. manifest.json stores a SHA-256 for each file. It is the hash of the decompressed JSONL bytes, not of the .gz file (verified for all 44 files at this revision). The table below lists them. scripts/r2_history.py also derives a suite identity from the file hashes of all 44 files.
• Decontamination. These rows are evaluation-only. The training builder (udm.data.build) adds the text of every test row to a MinHash near-duplicate index (word 5-gram shingles, 64 permutations in 16 bands, Jaccard threshold 0.8) and drops any training row whose state matches it. The R2 training-data policy keeps exact and MinHash decontamination against these suites and keeps complete groups together. Anything that is a near-duplicate of a row here must stay out of training data. Private AIV and DeepEdu evaluation records must never be training input. Sources that need to stay out of training because they feed these tracks include MS MARCO (TREC DL), ViNLI (NLI and ViNLI reranking) and the aiMy144 verifier test split.

File hashes (SHA-256 of decompressed content)

File	sha256
test/content_moderation/vihsd.jsonl.gz	d61222369a0d102836723e4a5efb52f725d77de250f5f9a296a149848a58d997
test/content_moderation/vihos.jsonl.gz	d4d6f4fccb93e55814bd84edbb5c71e51847c316e9d1c0a3bb297af8e2b4ef6f
test/content_moderation/victsd.jsonl.gz	91974d01d258ffab4653f9ca221005b99e3cd274bd1b5191af92fa0ccdf323e6
test/education/deepedubench-qa.jsonl.gz	674630208986e2aaffb9be625d084bb354895696ef0cc2856bca5de230c99e57
test/education/deepedubench-cdpk.jsonl.gz	2da4cf16e35ff67513c3a0116809510f84bcbe0b9f2b53fde8fcd926791f6204
test/education/deepedubench-grading-binary.jsonl.gz	40590ca3b2d3887116316174fc7e24b0e071f0f6aaeac70c9733def68597ea8c
test/education/deepedubench-grading-score.jsonl.gz	66058965b1b6f10bf598a66e8e5f806016a4d6b7d67210fd981d04645587c22a
test/fact_checking/aimy144-fact-checking-vi.jsonl.gz	cf595ca6ebd41517d28ab801c115dff53ca53d4f01b65e205b750b0bde37e63d
test/intent/massive-vi-vn.jsonl.gz	9d4f6e9c57fd18dc6d65b8777aee16c5c29420ced81503a370f07d03639ca2c1
test/model_routing/routerbench.jsonl.gz	62984f07799b8573cb7ea1b3598e38ae093aee2ed66b3c2bfb5a61b512cca436
test/multiple_choice/aiv-vlmu-mcqa-extend400.jsonl.gz	ed4bec07ba0e88a5664d7d422924038e11dfa08f75f384376a79fa7366901c2a
test/multiple_choice/aiv-hle-mc-vi.jsonl.gz	e3619b2138b58422bc4073965f9d4055742a2a3505b0123782eddbf1c840d337
test/multiple_choice/aiv-culture100.jsonl.gz	048e2eb76cafd8a3be82f80be371470afebb5254a2a0721039afbd0a31e119c1
test/multiple_choice/aiv-pragmatic50.jsonl.gz	f638ebd86fb99f1ddbe33653c2b7955f92645fb77935706f883307225b2acda7
test/multiple_choice/aiv-syntax50.jsonl.gz	2966651c474af4e93f9d6abaf5f13dd2eed980185d0bf3b5e31e5f06e4a27e4d
test/multiple_choice/mmlu-prox-vi.jsonl.gz	c5c0537c204a6ea2a55be5886937038e3f8c483b5eb98b9cda81b5a808f5e29c
test/multiple_choice/vimmrc-vi.jsonl.gz	47917f104451f63435e769bb7be7867f4d5140c08370e35e808bfcf6f124b8a8
test/nli/aiv-nli500.jsonl.gz	99ed521162a938d60c1a3926eb07007f65c7c256dc93535c5cf231e01d58e5d3
test/passage_reranking/zalo-legal.jsonl.gz	f50b1725e4b79f4fb0c5f9c05ee1cd3ff2ef205e1d9680c7e7918291c9ea76f4
test/passage_reranking/trec-dl19.jsonl.gz	96f1161588dfe83cfe2fa2cd4b0087e138637d7b86f00d9f6a657927b9c900df
test/passage_reranking/trec-dl20.jsonl.gz	6ebbdabe1523900177b1a72fae07a9bca45d4c62350231318cf8a9d34b193221
test/passage_reranking/wixqa-expertwritten.jsonl.gz	46195392e4ccc6c0a4a88d80b41c8785f05a9e5fcb9f9dc9a064dbf3e456bd44
test/passage_reranking/wixqa-simulated.jsonl.gz	4dd5a3210e4fa26841bd6f1b2ad4afce60ecad23a03c32a74ccba8d3afcd498b
test/passage_reranking/viequad-retrieval.jsonl.gz	344ac97d231267350eb9ef50e2e820dbbd49ab06d1ba38dbc1027ecebd23118f
test/passage_reranking/vinli-reranking.jsonl.gz	19485c5d90a13db2ad13d39d5666890e98c50c1868d3f2725f4e82d2fa5371c9
test/passage_reranking/table-markdown-vn.jsonl.gz	cf61c60bfeb9445492f25e691aed78bd900e3640cc4318031a2b32d8aa9ab93d
test/passage_reranking/tvpl-retrieval-vn.jsonl.gz	4ad209bf4af152a302b6e0be971f97a1a89664783652339fa0f889d3cce136ac
test/passage_reranking/askubuntu-reranking-vn.jsonl.gz	233ef445a041908ac2613b35e01aed345cc4b33f3f0eb784c9027ef0a807fff1
test/passage_reranking/stackoverflow-reranking-vn.jsonl.gz	c4f85a9aecf41761dc2ee0d9e07462432c1ea9f4dd291298dfd8ae7ffefc170f
test/passage_reranking/scidocs-reranking-vn.jsonl.gz	3848ae127f2a9940f0bb0009a55d54b065a0ad13a53b0ec71b58d2a3bff515bc
test/quality_scoring/fineweb-c-vi.jsonl.gz	53aed8850503f60cb8284f559e75258e5cdca2c793ebf8569c7376804aa1273a
test/safety_guardrail/deepedu.jsonl.gz	3087723c04f654b09ac0120b6b09e64797727e51e4df07857f74d3490588b59c
test/safety_guardrail/visafe.jsonl.gz	acd5e368f77d02b54adc94382707e596f63aedda13884fb4615a916a329abfc8
test/safety_guardrail/budguardrail.jsonl.gz	6bc86750873e213d04a04c481b55c55abccba337ff69e8a66379f9a59d7253a1
test/safety_guardrail/respan.jsonl.gz	ba77ac6b142cb009db9cb8ea540f5d4c71a2cff1a305b2d1a635d3ed147ae9f1
test/safety_guardrail/aiv-jbb.jsonl.gz	9f56fdf53c0bbf73f869ebf2b1409e797c0788bc5d1c1e4fddbe3a04e801d5dd
test/safety_guardrail/seasafeguard.jsonl.gz	b710918c2693a720ec7ee93fa9b672b86b52af18dbba865495b7c0bbc832b8d7
test/safety_guardrail/multijail.jsonl.gz	941b30ed203f20e7cbb502d89850b055a6b1e04fdbf683f1c82897a6179b88ca
test/sentiment/aiv-sa24-easy.jsonl.gz	15a7ff0d877769c967a0add9c4c82d4baa339c4627f4dd8994c263e8c6c07606
test/sentiment/aiv-sa26-hard.jsonl.gz	13f153f2c63b487655841ec6ee5103f4e82660887600ba637edf888e50b35a87
test/skill_routing/skillgym-sft.jsonl.gz	1ac6cce381a9d5ace1f4246d1e854f83ce0ecf99a4f97a8671512ae0d849964c
test/tool_selection/nemotron-tools.jsonl.gz	abed4398046a24e24f741c880a7807b5d041387adb9195a05cba642b5f3e98b6
test/tool_selection/toucan-kimi-k2.jsonl.gz	6b012c71abeb717fcc60740a19c517c9b78d1de7c6f6494046b46d2b8c60a049
test/tool_selection/deepseek-v4-pro.jsonl.gz	9bdd4ccdf3e4e1f8d2c5375249c928e9e53a2f225a9bc6ca600e6576e5cafd9e

Licence

The card-level licence is other. The records are copies or adaptations of third-party data, and each keeps the licence of its source, shown per track above. Several are unknown or unspecified (ViHSD, ViCTSD, aiMy144 fact checking, ViNLI reranking, TREC DL, the AIV and DeepEdu tracks, deepedu safety, SEASafeguard, SkillGym, DeepSeek-V4-Pro trajectories, RouterBench), mixed-source-licenses applies to budguardrail and respan, and some are non-commercial (ViMMRC CC-BY-NC-ND-4.0; the three translated GreenNode reranking sets CC-BY-NC-SA-4.0). Check each source before any use beyond evaluating a model, and before making this dataset public. The authors of this suite add no licence of their own.

Limitations

• No chance correction and unweighted track means: small tracks (24 to 100 rows) are noisy and count as much as large ones.
• Several tracks are translated or synthetic (see Origin), so they measure Vietnamese ability less directly than native ones.
• Single-gold reranking tracks (viequad-retrieval, tvpl-retrieval-vn, zalo-legal, table-markdown-vn, WixQA) label one passage per query, so useful unlabelled passages count against the model in nDCG; they have no keep-set metrics.
• Labels are those of the sources; label quality is not audited here (for example, aimy144-fact-checking-vi is marked provisional in the repo docs because its labelling model is unknown).