MaternaQA-es is distributed as three JSONL variants that share the same underlying content but expose it in different structures. The two SFT variants —Documentation Index
Fetch the complete documentation index at: https://mintlify.com/NicolasHoyosDevss/MaternaQA-es/llms.txt
Use this file to discover all available pages before exploring further.
sft_grounded and sft_closed_book — use the standard chat messages array format, making them directly compatible with training frameworks that accept conversational data. The qa_flat_jsonl variant stores every field as a top-level key, giving full access to provenance and quality metadata without unpacking nested objects. All three variants omit the operational judge quality metric fields (faithfulness, answer_relevancy, roundtrip_consistency, quality_verdict, quality_reason) that appear in the raw generation artifacts.
SFT Variants Schema
Bothsft_grounded and sft_closed_book records share this top-level structure.
Top-Level Fields
Ordered list of chat turn objects forming the conversation. Each record contains exactly three turns:
system, user, and assistant. In sft_grounded, the user turn includes the source context followed by the question. In sft_closed_book, the user turn contains only the question.Per-record audit and provenance metadata. Fields inside this object are consistent across both SFT variants.
Messages Turn Fields
Role of the speaker in the conversation. One of
system, user, or assistant.Text content of the turn. For the
system role, this is a fixed Spanish-language clinical instruction prompt. For user in sft_grounded, this is Contexto fuente: followed by the chunk text and then the question. For user in sft_closed_book, this is only the question. For assistant, this is the clinical answer.Metadata Fields
Stable identifier for the source corpus chunk. Format:
<pdf_slug>_<chunk_index> (e.g., s0213005x09004418_00006). Use this to join Q&A records back to the LM corpus dataset.Filename of the source clinical PDF from which the chunk was extracted (e.g.,
S0213005X09004418.pdf).Unique identifier for this specific Q&A pair. Format:
<chunk_id>_qa_<index> (e.g., s0213005x09004418_00006_qa_001).List of integer page numbers within the source PDF that the chunk spans.
Question type. One of
factual, definicion, comparacion, razonamiento, aplicacion, or hipotetico.Difficulty level assigned during generation. One of
basico, intermedio, or avanzado.List of clinical topic tags annotated for the source chunk (e.g.,
["newborn_care", "prenatal_care"]).Estimated token count of the source context chunk.
Quality score assigned to the chunk during corpus construction. Higher values indicate more clinically dense and useful content.
The source chunk text that grounds the Q&A pair. Present in both
sft_grounded and sft_closed_book metadata records. In sft_grounded, this same text is also embedded in the user message turn; in sft_closed_book, it is carried in metadata only and not shown to the model at inference time.Dataset split this record belongs to. One of
train, validation, or test.Complete sft_grounded Example
Complete sft_closed_book Example
qa_flat_jsonl Schema
The flat variant exposes all fields at the top level. The question and answer fields use Spanish-language names consistent with the generation pipeline.The clinical question in Spanish.
The clinical answer in Spanish, grounded in the source chunk.
The source chunk text used during grounded generation. This is the same content that appears in the
user turn of sft_grounded records.Stable chunk identifier in the format
<pdf_slug>_<chunk_index>. Links back to the LM corpus dataset.Unique identifier for the Q&A pair. Format:
<chunk_id>_qa_<index>.Filename of the source clinical PDF (e.g.,
S0213005X09004418.pdf).Integer page numbers within the source PDF that the chunk covers.
Question type. One of
factual, definicion, comparacion, razonamiento, aplicacion, or hipotetico.Difficulty level. One of
basico, intermedio, or avanzado.List of clinical topic tags (e.g.,
["prenatal_care", "infection"]).Estimated token count of the source context chunk.
Quality score for the source chunk from corpus construction. Higher values indicate more clinically relevant content.
Dataset split. One of
train, validation, or test.Complete qa_flat_jsonl Example
Quality Metric Fields
The fields
faithfulness, answer_relevancy, roundtrip_consistency, quality_verdict, and quality_reason are omitted from all publication files. They appear as null in the raw generation artifacts at datasets/obstetrics/qa/final/<split>/raw.jsonl and are produced by the operational judge used during data construction. They are not part of the formal Ragas evaluation results. See the Quality page for the formal evaluation methodology and scores.