Somtharu181coder/Annex_7_Educational_Indicators_2024-2025

Grounded Educational Indicators — Nepali SFT Dataset

File: grounded_education_indicators_nepali_sft.jsonl Records: 10,091 Language: Nepali (ne / ISO 639-3 npi), Devanagari script License: Apache-2.0 (permissive) Format: JSON Lines, ShareGPT-style conversational (human / gpt turns) Task type: Grounded question answering over structured (tabular) educational statistics Size on disk: ~15.5 MB

---

1. Overview

This dataset is a supervised fine-tuning (SFT) corpus of 10,091 Nepali-language question–answer pairs generated from a single structured source: an Excel workbook of national educational indicators for Nepal, titled "Annex 7 Educational Indicators, 2024–2025."

Each example asks a natural-language question in Nepali about a specific educational statistic (enrolment rates, survival/retention rates, out-of-school children, etc.) for a specific administrative unit, and pairs it with a factual, numerically grounded answer derived directly from the source spreadsheet. The dataset is designed to train or evaluate a model's ability to perform grounded retrieval and light quantitative reasoning (comparison, ranking, gap calculation, thresholding, synthesis) over tabular government statistics, in Nepali.

All answers are traceable to a specific row/sheet of the source workbook via the source_row_id and metadata_json fields, making this a fully attributable, non-hallucinated QA dataset by construction (subject to the caveats in Section 6 and the accompanying audit report).

---

2. Data Instance Format

Each line is a standalone JSON object. Example (formatted for readability):

{
  "id": "edu_np_local_direct_000001",
  "conversations": [
    {
      "from": "human",
      "value": "भोजपुर जिल्लाको आमचोक गाउँपालिकासम्बन्धी तथ्याङ्कबाट छात्राको कक्षा ९–१२ को खुद भर्ना दर बताउनुहोस्।"
    },
    {
      "from": "gpt",
      "value": "उक्त तथ्याङ्कमा भोजपुर जिल्लाको आमचोक गाउँपालिकाको छात्राको कक्षा ९–१२ को खुद भर्ना दर ५१.९२% रहेको छ।"
    }
  ],
  "source": "Annex 7 Educational Indicators, 2024-2025",
  "source_name": "educational_indicators_2024_2025",
  "source_repo": "local_excel_workbook",
  "source_config": "default",
  "source_split": "data",
  "source_revision": "2024-2025",
  "source_row_id": "excel_row:4",
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "license": "Apache-2.0",
  "license_tier": "permissive",
  "task_type": "grounded-question-answering",
  "generation_type": "factual",
  "condition": "grounded,structured-data",
  "url": "",
  "metadata_json": "{\"source_type\": \"Excel\", \"source_file\": \"...\", \"source_sheet\": \"first_sheet\", \"behavior\": \"direct_retrieval\", \"category\": \"direct_retrieval\"}",
  "category": "direct_retrieval",
  "behavior": "direct_retrieval"
}

Field reference

Field	Type	Description
id	string	Unique record identifier. Prefix encodes aggregation level and reasoning category (see §4).
conversations	array[2]	Exactly one human turn (question) followed by one gpt turn (answer).
source / source_name	string	Human-readable and machine-readable name of the underlying document (constant across the dataset).
source_repo	string	Origin type — always local_excel_workbook.
source_config, source_split	string	Fixed dataset-builder metadata (default, data).
source_revision	string	Fiscal-year edition of the source, 2024-2025.
source_row_id	string	Pointer to the originating Excel row (e.g. excel_row:4).
language, language_code, script	string	Always Nepali / npi / Deva.
license, license_tier	string	Always Apache-2.0 / permissive.
task_type	string	Always grounded-question-answering.
generation_type	string	Always factual (no opinion/creative generation).
condition	string	Always grounded,structured-data.
url	string	Empty for every record (no source URL captured).
metadata_json	string (JSON)	Nested metadata: source file name, sheet, and the behavior/category label duplicated from the top level.
category / behavior	string	Identical values; the reasoning skill each example targets (see §4).

---

3. Geographic Coverage & Aggregation Levels

Nepal's statistical hierarchy is National → Province (7) → District (77) → Local Level (753 municipalities/rural municipalities, i.e. gaunpalika and nagarpalika). The dataset generates questions at four of these levels:

Level	Records	Share	Example
Local (municipality/rural municipality)	9,031	89.5%	"भोजपुर जिल्लाको आमचोक गाउँपालिका..."
District (aggregated across local units)	903	8.9%	"ताप्लेजुङ जिल्लामध्ये ... सबैभन्दा बढी कहाँ छ?"
Province	122	1.2%	"कोशी प्रदेशमा जम्माको कक्षा १–५ को कुल भर्ना दर कति हो?"
National (all of Nepal)	35	0.3%	"सम्पूर्ण नेपालमा... विद्यालय बाहिर रहेका बालबालिकाको संख्या..."

Municipality/place-name mentions were detected across 76 of Nepal's 77 districts and 724 distinct local-level units, indicating broad (near-exhaustive) geographic coverage of the source workbook rather than a small illustrative sample. See the audit report for an important caveat regarding the spelling fidelity of local-level names.

---

4. Reasoning Categories (category / behavior)

The dataset is not a flat list of lookups — it spans 13 distinct question archetypes, ranging from direct fact retrieval to multi-step quantitative reasoning:

Category	Count	Share	What it tests
direct_retrieval	2,712	26.9%	Single-value lookup for one indicator, one place.
gender_comparison	840	8.3%	Compare male vs. female rates for the same indicator; identify which is higher and by how much.
ger_ner_gap	753	7.5%	Compute the percentage-point gap between Gross and Net Enrolment Rate.
oosc_distribution	753	7.5%	Report out-of-school-children counts, broken down by gender.
cross_level_comparison	753	7.5%	Compare an indicator across basic (grades 1–8) vs. secondary (9–12) education levels.
multi_indicator_synthesis	753	7.5%	Summarize three indicators for one place in a single answer.
evidence_based_summary	753	7.5%	Draw a conclusion by comparing two related indicators.
survival_progression	752	7.5%	Compare student survival/retention rate across grades 5, 8, and 10.
higher_indicator	751	7.4%	Identify which of two indicators has the higher value.
lower_indicator	751	7.4%	Identify which of two indicators has the lower value.
highest_ranking	185	1.8%	Find the local unit with the maximum value for an indicator, within a district.
lowest_ranking	185	1.8%	Find the local unit with the minimum value for an indicator, within a district.
threshold_identification	150	1.5%	List all local units in a district meeting a percentage threshold.

category and behavior are always identical in this release (redundant fields).

Indicators represented

Questions and answers draw on a consistent set of official Nepali education indicators, including:

• कुल भर्ना दर (Gross Enrolment Rate, GER) and खुद भर्ना दर (Net Enrolment Rate, NER), broken out by grade band (Grade 1 entrant rate, 1–5, 6–8, 9–10, 9–12, 11–12, and combined 1–8)
• प्रारम्भिक बाल विकास तथा पूर्वप्राथमिक कक्षा (early childhood development / pre-primary) enrolment
• दिगोपन दर (student survival/retention rate) to grades 5, 8, and 10
• विद्यालय बाहिर बालबालिका (out-of-school children), by sex

---

5. Dataset Statistics

• Total examples: 10,091 (all single-turn Q&A pairs; every conversations array has exactly 2 turns)
• Unique IDs: 10,091 / 10,091 (no duplicates)
• Exact duplicate question/answer pairs: 0
• Human turn length: 60–222 characters (mean 103, median 102)
• GPT turn length: 62–499 characters (mean 119, median 116)
• Answers containing a numeric percentage figure: 9,041 / 10,091 (89.6%) — the remainder are count-based (oosc_distribution), name-list (threshold_identification), or comparative-only answers
• Malformed / empty rows, JSON parse errors: 0
• Character encoding: consistent UTF-8, no replacement characters, all text Unicode-NFC normalized
• Numeral system: 100% Devanagari digits (०-९) in answers — no mixed Latin/Devanagari numerals

---

6. Known Limitations (see AUDIT_REPORT.md for full detail)

1. Local-unit name corruption. Across all 724 distinct municipality/rural-municipality names detected in the dataset, none contain a conjunct/halant character (्), while ordinary Nepali vocabulary in the same records does. This indicates the place names were corrupted upstream (most likely during extraction from the source Excel file, e.g. a legacy-font-to-Unicode mapping issue), producing misspelled place names such as "एकदर गाउँपालिका" where the correct form is likely "एकडारा गाउँपालिका." Numeric values and indicator terminology are not affected — only proper place names. Downstream users should treat place-name strings as approximate and, if exact spelling matters, cross-reference against an official gazette of Nepal's 753 local levels before deployment.
2. Single source, single time period. All records derive from one workbook covering FY 2024–2025. There is no temporal variation (no year-over-year trend questions) and no cross-source verification.
3. Template-visible generation. Answers closely mirror question phrasing (by design, for grounded QA), so a model trained solely on this data may over-fit to this phrasing style; consider mixing with paraphrase-diverse data for robustness.
4. No negative/unanswerable examples. Every question has a directly retrievable answer in the source table; the dataset does not teach a model to recognize when data is missing or a question is out of scope.
5. Coverage skew. 89.5% of records are at the local (municipality) level, so evaluation slices at the province/national level are comparatively small (122 and 35 records respectively).

---

7. Suggested Uses

• Fine-tuning or evaluating Nepali-language LLMs on grounded, numeric, structured-data QA
• Benchmarking table-to-text / retrieval-augmented generation pipelines in a low-resource language
• Studying multi-step quantitative reasoning (gap computation, ranking, thresholding) in Nepali
• As a component of a larger multilingual grounded-QA mixture

8. Suggested Citation / Provenance Note

When redistributing, retain the original provenance metadata (source, source_name, source_revision, source_row_id) so that any downstream consumer can trace an answer back to "Annex 7 Educational Indicators, 2024–2025," row/sheet level.

---

This README was generated from a direct structural and statistical audit of the dataset file — see AUDIT_REPORT.md for the full methodology and findings.