Ephraimmm/Clinical_NLP_benchmark_dataset

Synthetic African Clinical NLP Benchmark (8K)

Overview

This is a fully synthetic benchmark of 8,000 short "patient query -> assistant response" pairs styled after healthcare interactions in African countries. Each record pairs a symptom-style query with a templated clinical-style response and metadata (condition category, medical specialty, urgency level, and an assigned hallucination-risk label). It is designed as a lightweight testbed for evaluating clinical/medical LLMs and chatbots on symptom triage, response classification, and safety-oriented behavior (e.g. avoiding overconfident diagnostic claims) — pairing well with generative clinical models such as SOAP-note generators, which need to be evaluated on how they handle patient-style input.

Dataset Structure

• Total records: 8,000 (verified by counting rows in the JSONL file)
• File: synthetic_african_clinical_nlp_benchmark_8k.jsonl (single split, no train/test division provided)

Each row is a JSON object with the following fields:

Field	Description
id	Integer row identifier
country	One of 7 countries: Kenya, Nigeria, Ghana, Tanzania, Rwanda, South Africa, Uganda
language	One of 4 values: English, Swahili, Yoruba, Nigerian Pidgin
medical_specialty	One of 7 categories: General Practice, Pediatrics, Cardiology, Neurology, Dermatology, Infectious Disease, Emergency Medicine
condition_category	One of 7 categories: Malaria, Typhoid, Hypertension, Diabetes, Asthma, Migraine, Food Poisoning
patient_query	A short, templated symptom description ending in a question, e.g. "I am experiencing fever and chills. I live in Nigeria. What could this indicate?"
assistant_response	A short, templated reply naming the likely condition and a generic recommendation, always ending with a synthetic-data disclaimer
urgency_level	low, medium, or high
hallucination_risk	low, medium, or high (an assigned label, not a measured/model-derived score)
synthetic	Always true for every row
source	Always "synthetic_generation_pipeline_v1" for every row

Data Format

Newline-delimited JSON (JSONL), one record per line, UTF-8 encoded. No image, audio, or binary data.

Intended Use

• Smoke-testing and lightweight benchmarking of clinical/medical LLMs and chatbots on symptom-to-response tasks
• Classification-style evaluation (e.g. predicting condition_category, urgency_level, or hallucination_risk from patient_query)
• Retrieval-augmented generation (RAG) and hallucination-detection pipeline testing on a controlled, low-risk synthetic corpus
• Educational and research use in clinical NLP, particularly work with an African healthcare framing (multiple countries and languages represented)

Limitations

• Fully synthetic, template-generated data. Verified inspection shows the dataset was produced from a small set of templates: only 294 unique patient_query strings and 35 unique assistant_response strings occur across all 8,000 rows. This makes the benchmark useful for quick, repeatable sanity checks but not representative of the linguistic diversity of real clinical text.
• Internal inconsistency in the country field. In a majority of rows (roughly 6,867 of 8,000, ~86%), the country named inside the patient_query text ("I live in ...") does not match the row's country field. This appears to be a template-combination artifact and should be accounted for (or filtered) before using country as a reliable label.
• No real patient data. The dataset contains no real patients, clinicians, or clinical records, and no de-identification was needed or performed — everything is generated.
• hallucination_risk is an assigned category, not a measured outcome. It is not derived from evaluating any model and should not be interpreted as ground-truth model-hallucination data.
• Not clinically validated. Content has not been reviewed by medical professionals for accuracy and must not be used for real diagnosis, triage, or any clinical decision-making.
• No official train/test/validation split is provided; users must create their own splits.
• Limited scope. Only 7 condition categories, 7 specialties, 4 languages, and 7 countries are represented, so this is a narrow benchmark rather than a comprehensive one.

Author

Curated by Ephraimmm