SAUDADE-SFT — Conversational Instruction-Tuning Corpus (European Portuguese)

SAUDADE-SFT is the large-scale synthetic supervised fine-tuning corpus that accompanies the SAUDADE temporal reasoning benchmark, released as part of the AMALIA project.

292,338 multi-turn dialogues across seven conversational personas, built from verified temporal facts extracted from Portuguese Wikipedia, in three multi-turn variants plus the single-turn instruction-following partitions.

Contents

Variant	Files	Items	Notes
Tiser	sft_multi_turn/tiser/saudade_sft_tiser_relation_with_distractors.jsonl sft_multi_turn/tiser/saudade_sft_tiser_fact_retrieval.jsonl	149,231 + 143,107 = 292,338	structured reasoning blocks (<reasoning>, <timeline>, <reflection>, <answer>)
Clean	sft_multi_turn/clean/saudade_sft_clean_relation_with_distractors.jsonl sft_multi_turn/clean/saudade_sft_clean_fact_retrieval.jsonl	149,176 + 143,082 = 292,258	same dialogues without the reasoning blocks
Ready to train	sft_multi_turn/ready_to_train/saudade_sft_train.jsonl sft_multi_turn/ready_to_train/saudade_sft_val.jsonl	290,797 + 1,461 = 292,258	the same Clean dialogues in messages format (user/assistant), split exactly as used for supervised fine-tuning (validation never trained on)
Single-turn (ShareGPT)	sft_single_turn/saudade_sft_single_turn_fact_retrieval.jsonl	161,329	Fact Retrieval, human/gpt pairs
Single-turn (glue)	sft_single_turn/saudade_sft_single_turn_relation_with_distractors.jsonl	179,554	Relation with Distractors, glue format

Loading

Each variant is a separate configuration. For the multi-turn variants the splits are the task partitions; ready_to_train exposes the train/validation split used for supervised fine-tuning. The two single-turn partitions have different schemas, so each is its own configuration. multi_turn_tiser is the default configuration.

from datasets import load_dataset

repo = "amalia-llm/SAUDADE-SFT"

tiser_fr  = load_dataset(repo, "multi_turn_tiser", split="fact_retrieval")
tiser_rwd = load_dataset(repo, "multi_turn_tiser", split="relation_with_distractors")

clean_fr  = load_dataset(repo, "multi_turn_clean", split="fact_retrieval")

train = load_dataset(repo, "ready_to_train", split="train")
val   = load_dataset(repo, "ready_to_train", split="validation")

single_fr  = load_dataset(repo, "single_turn_fact_retrieval", split="train")
single_rwd = load_dataset(repo, "single_turn_relation_with_distractors", split="train")

Counts

The corpus as generated contains 292,342 dialogues; the released variants exclude the dialogues with an empty or structurally invalid answer (4 in the Tiser variant, 84 in the Clean variant), which is why the two variants have 292,338 and 292,258 dialogues respectively.

Personas

leigo, jornalista, cético, estudante, entrevistador, curioso, historiador — seven pragmatic registers with markedly different dialogue styles.

Schema (multi-turn)

Field	Description
entity	source Wikipedia entity
persona	pragmatic register of the dialogue
n_tiser, n_casual	number of structured-reasoning / casual turns (Tiser variant)
conversations	turns as role (human/gpt) and content — Tiser / Clean
messages	turns as role (user/assistant) and content — ready-to-train variant

Field names (mapping)

The multi-turn Tiser and Clean files keep the pipeline's original dataset field; the ready_to_train export contains only messages, and the single-turn files use domain and category instead.

dataset in the JSONL	Task name in the paper / thesis
awareness	Temporal Relation with Distractors
cognition	Temporal Fact Retrieval
reasoning	Binary Temporal Ordering (dialogues; not part of this corpus)

License

Apache-2.0 (see LICENSE).

Citation

@inproceedings{simplicio2026amalia,
  title     = {{AMALIA}: A Fully Open Source Large Language Model for European Portuguese},
  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 Vieira, In{\^{e}}s and Guerra, Rui and Furtado, James and
               Canaverde, Beatriz and Paulo, Iago and Ramos, Vasco and Gl{\'{o}}ria-Silva, Diogo and
               Faria, Miguel and Treviso, Marcos and Gomes, Daniel and Gomes, Pedro and
               Semedo, David and Martins, Andr{\'{e}} and Magalh{\~{a}}es, Jo{\~{a}}o},
  booktitle = {arXiv preprint arXiv:2603.26511},
  year      = {2026}
}

The SAUDADE benchmark and SFT resource are described in the accompanying ECIR 2027 resource paper.