TigreGotico/jevasio-decisions-v2

Jevásio decisions v2

Content warning. The safety sources contain harmful requests (violence, self-harm, hate, illegal activity) and toxic messages. They are here so that a model can learn to detect them.

This dataset holds typed decisions in Portuguese, most of them European Portuguese (pt-PT), for training Jevásio, a decision model for European Portuguese. A record is a state (text or JSON) and one or more questions. Each question is a choice (pick one option), a score (an ordered level) or a noul (yes or no), with its label. The format is the training format of Kev, an open model that answers this request shape with calibrated probabilities in one forward pass.

The labels of 63 sources come from the source data: an answer key, a safety annotation, a preference pair, a variety tag, or a wrong option taken from another row. The labels of the tierc rows come from a teacher model, AMALIA-9B (see Teacher-labelled rows). The source data is the amalia-llm collection of the AMALIA project.

Relationship to the first version

This version contains all of TigreGotico/jevasio-decisions: 147,254 of its 147,254 training states are here, with the same splits. It has 335,063 training requests against 147,254, 2.3 times as many, from 36 new sources. The additions are mostly new kinds of decision rather than more of the same: European against Brazilian Portuguese, assistant intents, entailment and similarity, news, legal and political classification, offence and sentiment, and harder faithfulness checks. Some of the new data has silver labels (the variety corpus), some is machine-translated (extraGLUE), and some is generated (see Edited negatives and generated texts).

Files

File	Requests	Content
train.jsonl	335,063	training requests
dev.jsonl	35,984	model selection
calib.jsonl	28,315	temperature fitting
test.jsonl	37,972	held-out evaluation, including the evaluation-only sources
manifest.json		per source: dataset revision, licence, rows read, requests per split, and each drop reason with its count
tierc/unlabelled.jsonl	11,197 states	the teacher-labelled questions before labelling, with every question type
tierc/teacher-amalia.jsonl	33,591 questions	the teacher's probability over the options of every question

The state is a string in some records and a JSON object or list in others, as in the Jev request format. One table schema cannot hold that column, so the Hub viewer is off. Read the files as JSON lines:

import json
from huggingface_hub import hf_hub_download

path = hf_hub_download("TigreGotico/jevasio-decisions-v2", "train.jsonl", repo_type="dataset")
records = [json.loads(line) for line in open(path, encoding="utf-8")]

To train Kev on it, run kev.train --data train.jsonl --init_from jaredpalmer/kev-0.8b --p_none 0 --p_none_distract 0 --p_distract 0. The three zeros turn off Kev's English "none of the above" options and English distractors.

Record format

{"state": {"question": "..."},
 "questions": {"answer": {"type": "choice", "instructions": "Escolhe a resposta correta.",
                            "criteria": {"opt_1": "...", "opt_2": "..."}, "label": "opt_2", "src": "csqa_pt"}},
 "_meta": {"source": "csqa_pt", "repo": "amalia-llm/commonsense_qa-mt-pt", "revision": "...", "split": "train",
           "variety": "mt-pt", "id": "...", "group_id": "...", "text_sha256": "..."}}

Keys are code identifiers, so they are in English and only the values are in Portuguese, as in a request a developer writes: the state's field names, the instruction keys (question, focus), the names of choice options (health, finance) and the question ids.

The label is the option key for choice, true or false for noul, and the level index from 0 for score. Option order is shuffled per record, so the key never carries the answer. _meta.revision is the commit of the source dataset that the build read. A tierc record also carries _meta.teacher_probs, the teacher's probabilities over the options.

Splits

A record goes to train, dev, calib or test by a hash of its group key, weighted 80/10/5/5. Records that must stay together share a group key, for example the two sentences of a WinoGrande twin pair. Sources with their own splits keep them. When the upstream test labels are hidden (CommonsenseQA, PIQA, WinoGrande), the validation split is divided into dev and test. A training record is removed when its state appears in any held-out split of any source.

Each source gives at most 10,000 training requests and 1,000 requests to each held-out split. Evaluation-only sources give up to 5,000 test requests and nothing to train.

Sources

The variety column is pt-PT for native European Portuguese data, pt-BR for Brazilian data, and mt-pt for machine translations into Portuguese. The licence column gives the licence and where it is stated: the amalia-llm card, or the card or repository of the original dataset.

Source	Dataset	Variety	Train	Dev	Calib	Test	Licence
assin2	nilc-nlp/assin2	pt-BR	6,069	487	380	1,000	not stated (card: unknown)
assin_ptbr	nilc-nlp/assin	pt-BR	2,348	500	151	1,000	not stated (card: unknown)
assin_ptpt	nilc-nlp/assin	pt-PT	2,376	500	124	1,000	not stated (card: unknown)
bbh_pt	amalia-llm/bigbenchhard-mt-pt	mt-pt	2,189	285	123	144	MIT (upstream suzgunmirac/BIG-Bench-Hard GitHub)
boolq	PORTULAN/extraglue	mt-pt	7,943	1,000	484	868	MIT (Hub card)
cb	PORTULAN/extraglue	mt-pt	236	43	14	13	MIT (Hub card)
copa	PORTULAN/extraglue	mt-pt	375	67	25	33	MIT (Hub card)
corege_fos	amalia-llm/CorEGe-PT	pt-PT	10,000	1,000	1,000	1,000	CC-BY-NC-SA-4.0 (amalia-llm card)
csqa_pt	amalia-llm/commonsense_qa-mt-pt	mt-pt	9,004	777	528	421	MIT (upstream tau/commonsense_qa Hub card)
dpo_pt	amalia-llm/DPO-Dataset	pt-PT	10,000	1,000	1,000	1,000	not stated; mixed upstream sources, including CC-BY-NC-SA-4.0 persona data
dsl_tl	amalia-llm/dsl_tl_fewshot	pt-PT	2,274	298	124	147	not stated (no upstream licence found for DSL-TL)
egsciqa	amalia-llm/EGSciQA-ptPT-V1	pt-PT	10,000	1,000	772	1,000	not stated; generated from CorEGe-PT (CC-BY-NC-SA-4.0)
fortuna	johnpaulbin/portuguese-hate-speech	pt-BR	4,538	580	269	279	not stated (card: unknown)
frmt	amalia-llm/frmt_fewshot	pt-PT	196	24	10	4	CC-BY-SA-3.0 (upstream google-research FRMT README)
hatebr	franciellevargas/HateBR	pt-BR	5,620	675	361	344	Apache-2.0 (Hub card)
massive	SetFit/amazon_massive_intent_pt-PT	pt-PT	10,000	1,000	658	1,000	CC-BY-4.0 (upstream AmazonScience/massive)
mrpc	PORTULAN/extraglue	mt-pt	3,433	266	232	142	MIT (Hub card)
olid_br	dougtrajano/olid-br	pt-BR	4,378	565	270	1,000	CC-BY-4.0 (Hub card)
openbookqa_pt	amalia-llm/openbookqa-mt-pt	mt-pt	4,530	488	309	478	Apache-2.0 (upstream allenai/OpenBookQA GitHub)
parliament	luist18/portuguese-parliament-interventions	pt-PT	2,423	302	137	136	MIT (Hub card)
piqa_pt	amalia-llm/piqa-mt-pt	mt-pt	10,000	1,000	865	542	not stated (upstream ybisk/piqa Hub card says unknown)
porvid_journalistic	LCA-PORVID/portuguese_vid	pt-PT	9,681	968	975	826	not stated (silver labels by source corpus)
porvid_legal	LCA-PORVID/portuguese_vid	pt-PT	9,798	978	987	917	not stated (silver labels by source corpus)
porvid_literature	LCA-PORVID/portuguese_vid	pt-PT	10,000	1,000	1,000	1,000	not stated (silver labels by source corpus)
porvid_politics	LCA-PORVID/portuguese_vid	pt-PT	9,975	998	999	985	not stated (silver labels by source corpus)
porvid_social_media	LCA-PORVID/portuguese_vid	pt-PT	6,747	673	647	697	not stated (silver labels by source corpus)
porvid_web	LCA-PORVID/portuguese_vid	pt-PT	9,813	987	980	996	not stated (silver labels by source corpus)
pt_culture	amalia-llm/PT-Culture_Data	pt-PT	10,000	1,000	1,000	1,000	Apache-2.0 (amalia-llm card)
ptradutor	amalia-llm/amalia-PTradutor	pt-PT	10,000	1,000	995	987	not stated
ptradutor_hard	amalia-llm/amalia-PTradutor	pt-PT	5,000	489	493	491	not stated
publico	duarteocarmo/cc_news_publico	pt-PT	10,000	1,000	1,000	1,000	not stated (Público articles from Common Crawl News)
puntuguese	Superar/Puntuguese	pt-BR	3,737	569	246	1,000	CC-BY-SA-4.0 (Hub card)
rewrite_faithful	amalia-llm/smol-rewrite-PT	mt-pt	10,000	1,000	981	1,000	Apache-2.0 (upstream HuggingFaceTB/smoltalk card, smol-rewrite)
rewrite_faithful_hard	amalia-llm/smol-rewrite-PT	mt-pt	5,005	483	488	493	Apache-2.0 (upstream HuggingFaceTB/smoltalk card, smol-rewrite)
rte	PORTULAN/extraglue	mt-pt	2,340	175	149	101	MIT (Hub card)
saudade	amalia-llm/saudade-pt	pt-PT	6,804	854	457	458	not stated
siqa_pt	amalia-llm/siqa-mt-pt	mt-pt	10,000	1,000	1,000	685	not stated (no upstream licence found)
sst2	PORTULAN/extraglue	mt-pt	10,000	573	1,000	298	MIT (Hub card)
stj	MartimZanatti/Descriptors_STJ	pt-PT	10,000	1,000	1,000	1,000	MIT (Hub card)
stsb	PORTULAN/extraglue	mt-pt	5,370	976	335	512	MIT (Hub card)
summary_faithful	amalia-llm/smol_summarize_pt	mt-pt	10,000	1,000	1,000	994	Apache-2.0 (upstream HuggingFaceTB/smoltalk card, smol-summarize)
summary_faithful_hard	amalia-llm/smol_summarize_pt	mt-pt	4,911	477	494	508	Apache-2.0 (upstream HuggingFaceTB/smoltalk card, smol-summarize)
synth_offensive	generated	pt-PT	2,391	345	119	180	CC-BY-NC-SA-4.0 (generated for this dataset)
synth_sentiment	generated	pt-PT	2,468	329	105	130	CC-BY-NC-SA-4.0 (generated for this dataset)
synth_support_route	generated	pt-PT	2,040	286	126	139	CC-BY-NC-SA-4.0 (generated for this dataset)
synth_urgency	generated	pt-PT	1,593	206	101	102	CC-BY-NC-SA-4.0 (generated for this dataset)
synth_variety	generated	pt-PT + pt-BR	2,305	281	165	124	CC-BY-NC-SA-4.0 (generated for this dataset)
toxichat	amalia-llm/toxichat-ptpt	pt-PT	3,934	509	228	239	CC-BY-NC-4.0 (upstream lmsys/toxic-chat Hub card)
tupy	Silly-Machine/TuPyE-Dataset	pt-BR	10,000	1,000	1,000	1,000	CC-BY-4.0 (Hub card)
tweet_sentiment	cardiffnlp/tweet_sentiment_multilingual	pt-BR	1,728	323	111	869	not stated
wildguardmix	amalia-llm/wildguardmix-ptpt	pt-PT	1,363	174	86	102	ODC-BY (upstream allenai/wildguardmix Hub card)
winogrande_pt	amalia-llm/winogrande-mt-pt	mt-pt	10,000	826	1,000	396	Apache-2.0 (upstream allenai/winogrande GitHub)
wmt24pp	amalia-llm/wmt24pp_xx_to_pt_10k	pt-PT	8,008	1,000	463	497	Apache-2.0 (upstream google/wmt24pp Hub card)
wmt24pp_hard	amalia-llm/wmt24pp_xx_to_pt_10k	pt-PT	3,997	490	217	238	Apache-2.0 (upstream google/wmt24pp Hub card)
tierc	four amalia-llm conversation sets	pt-PT	8,953	1,158	532	554	not stated (persona_general, persona_nemotron, wikipedia_conversations); Apache-2.0 (smoltalk2_everyday_conv_pt)
alba_mcq	amalia-llm/alba_mcq	pt-PT	0	0	0	237	not stated
alba_rating	amalia-llm/alba	pt-PT	0	0	0	719	not stated
cultura_viva	amalia-llm/cultura-viva-pt-mcq	pt-PT	0	0	0	2,463	not stated
harmbench	amalia-llm/harmbench-ptpt	pt-PT	0	0	0	300	MIT (upstream walledai/HarmBench Hub card)
porvid_gold	LCA-PORVID/gold_labelled	pt-PT	0	0	0	219	not stated (manual labels)
pt_exams	amalia-llm/pt_exams	pt-PT	0	0	0	1,346	not stated
pt_text_completion	amalia-llm/pt_text_completion	pt-PT	0	0	0	69	not stated
sst	amalia-llm/sst_pt-pt	pt-PT	0	0	0	100	CC-BY-4.0 (amalia-llm card)
xstest	amalia-llm/xstest_ptpt	pt-PT	0	0	0	450	CC-BY-4.0 (amalia-llm card)

Brazilian data and the variety question

European and Brazilian Portuguese are one language. For most tasks the variety of the text does not matter, so the Brazilian sources (assin2, assin_ptbr, fortuna, hatebr, olid_br, puntuguese, tupy, tweet_sentiment) teach sentiment, offence and sentence relations as the European ones do. Only the variety question depends on the variety. Its 6 porvid sources come from the six domains of the LCA-PORVID corpus (56,014 training requests), in the proportions each domain has. The corpus labels a text by where it came from, so a text whose marker words contradict its label is dropped: spellings such as económico against econômico, and words such as telemóvel against celular. That check dropped 614 texts. In porvid_journalistic, porvid_legal, porvid_politics, porvid_social_media, porvid_web, claude-haiku-5-5 then labelled every text blind, and 5,234 texts whose label it did not confirm were dropped. On the 219 gold texts that judge agrees with the manual label 205 times. porvid_literature keeps its corpus labels: on classic Brazilian prose, which follows norms close to European ones, the judge called 6 of 15 Brazilian gold texts European. porvid_gold, the corpus's manually labelled set, is for evaluation only.

Edited negatives and generated texts

The _hard sources pair each faithful summary, rewrite or translation with a copy in which a model changed exactly one fact: a number, a date, a name, a place, who did what, or a negation. An edit is kept when it changes at most three words, or 15% of a longer text. The edit shares the split and group of its original, so the model sees both sides of the pair. 24,274 edits, made by claude-haiku-5-5, go-glm.

The synth_ sources are texts that one model wrote for a target label (offensive, sentiment, support_route, urgency, variety), in pt-PT and, for the variety task, also in pt-BR, each call with a drawn domain, register and length. A second model then labelled every text blind, from the options the question offers, and a text is kept only when the two labels agree. Texts of one call share a group. 13,535 texts kept; written by claude-sonnet-5-5, checked by claude-haiku-5-5.

Teacher-labelled rows

The tierc rows ask Jev-style questions about real pt-PT conversations from persona_general, persona_nemotron, wikipedia_conversations and smoltalk2_everyday_conv_pt. The state is the user's request, and in half of the records also the assistant's answer.

AMALIA-9B-1225-DPO labels each question. It reads the question with its options as letters, and the label is the letter with the highest next-token probability. No text is generated. These labels are the teacher's judgement, not a gold standard.

Question	Type	Questions kept
portugal	noul	5,822
registo	score	5,676
tema	choice	5,847
tipo_pedido	choice	5,761
variedade_resposta	choice	2,369

Two question types were dropped because the teacher gave nearly every record the same label: whether the text holds personal data (dados_pessoais, 5,843 questions) and whether the answer is complete (resposta_completa, 2,273 questions). A label that is the same for over 98% of records teaches a bias, not a decision.

Licence

This dataset is released under CC BY-NC-SA 4.0, for non-commercial research. Some sources are non-commercial (toxichat, CorEGe-PT and the data generated from it). CorEGe-PT is share-alike, and FRMT is CC BY-SA 3.0.

The source cards for 26 sources state no licence: alba_mcq, alba_rating, assin2, assin_ptbr, assin_ptpt, cultura_viva, dpo_pt, dsl_tl, egsciqa, fortuna, piqa_pt, porvid_gold, porvid_journalistic, porvid_legal, porvid_literature, porvid_politics, porvid_social_media, porvid_web, pt_exams, pt_text_completion, ptradutor, ptradutor_hard, publico, saudade, siqa_pt, tweet_sentiment. These records follow the terms under which the AMALIA project publishes its data. If you hold rights to one of these sources and want it removed, open a discussion on this repository.

Limits

• Six knowledge sets are machine translations. They teach the task format but carry translation artefacts. The native pt-PT sources (saudade, pt_culture, corege_fos, egsciqa, pt_exams, alba, cultura_viva) are the better test of pt-PT ability.
• The wrong options in pt_culture and the wrong outputs in the faithfulness sources come from other rows. They are easier than real errors.
• DPO rows are kept when Portuguese function words outnumber English ones three to one. This is a heuristic, not a language identifier.
• harmbench and sst contain harmful requests only, so they measure recall and nothing else.
• dsl_tl labels a text PT-PT or PT-BR. The source has no "both" label, so a text that fits both varieties still gets one label.
• One teacher labels the tierc rows, and nobody checked its answers. Its errors are in the labels.
• The porvid labels come from the source of each text, not from a reader. The marker check removes only the clear mistakes.
• The extraGLUE sources (rte, cb, mrpc, boolq, stsb, copa, sst2) are machine translations from English.
• The parliament text is lowercased, and some of its articles are missing.

Citation

The source data comes from the AMALIA project. Cite it, and the original dataset of each source you use:

@inproceedings{simplicio-etal-2026-amalia,
    title = "{AMALIA}: A Fully Open Large Language Model for {E}uropean {P}ortuguese",
    author = "Simpl{\'i}cio, Afonso and Vinagre, Gon{\c{c}}alo and Ramos, Miguel Moura and Tavares, Diogo and Ferreira, Rafael and Attanasio, Giuseppe and Alves, Duarte M. and Calvo, In{\^e}s and others",
    booktitle = "Proceedings of the 17th International Conference on Computational Processing of {P}ortuguese ({PROPOR} 2026) - Vol. 1",
    year = "2026",
    address = "Salvador, Brazil",
    publisher = "Association for Computational Linguistics",
    url = "https://aclanthology.org/2026.propor-1.38/",
    pages = "380--391"
}