README.md

---
language:
- ne
license: cc-by-4.0
task_categories:
- question-answering
- text-generation
tags:
- nepali
- devanagari
- grounded-qa
- education
- synthetic
- instruction-tuning
- sharegpt
size_categories:
- n<1K
---

# Nepali Grounded Education QA (OpenHermes-format)

A small, fact-grounded Nepali instruction-tuning dataset of question–answer pairs about
student enrollment statistics from Nepal's Ministry of Education. Every answer is
anchored to a real numeric value pulled from government open data — nothing in the
answers is model-hallucinated.

## Dataset Summary

| | |
|---|---|
| **Rows** | 611 |
| **Language** | Nepali (Devanagari script) |
| **Format** | ShareGPT / `hermes-instruction-response` (`conversations: [human, gpt]`) |
| **Domain** | Education statistics — janajati (indigenous/ethnic) student enrollment |
| **Source** | Nepal Ministry of Education, via [Open Data Nepal](https://opendatanepal.com) |
| **Generation method** | LLM-paraphrased Nepali question/answer pairs, grounded in a single numeric fact per row |
| **License** | CC BY 4.0 (see [Licensing](#licensing) for full reasoning) |

## Dataset Structure

Each line is a standalone JSON object:

```json
{
  "id": "grounded-qa-000001-Grade-1",
  "conversations": [
    {"from": "human", "value": "वर्ष 2011-012 मा ताप्लेजुङ जिल्लाका सबै प्रकारका विद्यालयहरूमा कक्षा 1 मा कति जनाजाति विद्यार्थीहरू भर्ना भएका थिए?"},
    {"from": "gpt", "value": "वर्ष 2011-012 मा ताप्लेजुङ जिल्लाका सबै प्रकारका विद्यालयहरूमा कक्षा 1 मा 9278 जनाजाति विद्यार्थीहरू भर्ना भएका थिए।"}
  ],
  "source_provenance": {
    "grounding": {"district": "Taplejung", "grade": "Grade 1", "value": "9278"},
    "metadata": {
      "dataset_name": "Number of Schools and Students FY 2003 to 2015 BS",
      "dataset_domain": "Ministry of Education",
      "resource_name": "Janajati enrollment at all levels by district in all types of School 2011-012 - Sheet1.csv",
      "license_id": "notspecified"
    },
    "behavior": "paraphrase"
  },
  "input_format": "hermes-instruction-response",
  "source_task_id": "grounded-qa-000001-Grade-1"
}
```

**Fields:**

- `id` / `source_task_id` — unique row identifier, encodes district and grade
- `conversations` — a single human/gpt turn pair, ready for SFT
- `source_provenance.grounding` — the exact district, grade, and numeric value the answer is anchored to
- `source_provenance.metadata` — source dataset name, domain, and originating CSV resource
- `behavior` — generation strategy used for this row (currently `paraphrase` for 100% of rows)

## Data Quality & Audit

All 611 rows were programmatically audited prior to release. Results below are from
the actual audit pass, not estimated.

### Structural integrity

| Check | Result |
|---|---|
| Rows with malformed/missing `conversations` | 0 / 611 |
| Rows with incorrect turn roles (not `human` → `gpt`) | 0 / 611 |
| Rows with empty question or answer text | 0 / 611 |
| Duplicate `id` values | 0 / 611 |
| Duplicate exact question+answer pairs | 0 / 611 |
| Rows missing a required top-level field (`id`, `conversations`, `source_provenance`, `input_format`, `source_task_id`) | 0 / 611 |

### Language & content integrity

| Check | Result |
|---|---|
| Suspicious/invisible Unicode characters (zero-width chars, stray symbols outside Devanagari + standard punctuation ranges) | 0 / 611 |
| Rows with detectable English prose contamination | 0 / 611 |
| Numeral encoding in answers | 100% Western/Arabic digits (0% Devanagari numerals — see [Limitations](#known-limitations)) |

### Factual grounding fidelity

| Check | Result |
|---|---|
| Rows where the grounded numeric value (`source_provenance.grounding.value`) appears verbatim in the generated answer | **611 / 611 (100%)** |

This is the most important check for a grounded-QA dataset: every answer contains the
exact figure it was supposed to report, with no numeric drift introduced during
paraphrasing/translation.

### Coverage

| | |
|---|---|
| Unique districts covered | 76 |
| Grade levels covered | Grade 1–12 |
| Grade distribution | Uneven — ranges from 18 rows (Grade 10) to 63 rows (Grade 4); see note below |
| Source CSV resources | 1 (`Janajati enrollment at all levels by district in all types of School 2011-012 - Sheet1.csv`) |
| Reporting year | 2011–012 BS only |

### Row length consistency

| | Human turn (chars) | GPT turn (chars) |
|---|---|---|
| Min | 109 | 108 |
| Mean | 114.5 | 115.1 |
| Max | 140 | 125 |

Narrow, consistent length range — indicates a stable question/answer template rather
than free-form generation drift.

## Known Limitations

- **Single topic, single year.** Every row is derived from one CSV (janajati enrollment
  by district and grade, fiscal year 2011–012 BS). This dataset is narrow by design —
  useful for grounded numeric QA in Nepali, not a general-purpose instruction dataset.
- **Uneven grade coverage.** Grade 10 has roughly a third as many rows as Grade 4,
  reflecting gaps in the source CSV rather than sampling choice.
- **Arabic numerals only.** All numeric values in answers use Western digits (0–9)
  rather than Devanagari numerals (०-९). If your downstream use case needs Devanagari
  numeral consistency, a numeral-conversion pass is required before use.
- **Single generation behavior.** All 611 rows use the `paraphrase` behavior — there is
  no lexical/structural diversity from alternate generation strategies.
- **Inherited source metadata quirk.** The `metadata.publisher` field on every row
  reads "Agriculture Inputs Company Limited," which does not match the actual
  publishing body (Ministry of Education). This is inherited as-is from Open Data
  Nepal's portal metadata and was not corrected, to preserve source-provenance
  traceability; it should not be read as the actual publisher of the underlying
  statistics.
- **No held-out human review.** Audit checks above are automated (structural,
  Unicode, contamination, and grounding-fidelity checks). No manual human-annotator
  review pass has been run on this release.

## Licensing

- The underlying enrollment statistics originate from Nepal's Ministry of Education
  and are distributed as open government data via
  [Open Data Nepal](https://opendatanepal.com). The source portal lists this specific
  resource's license as `notspecified` — the publisher did not declare a formal
  license, though it is published as open data.
- Bare statistics (counts, figures) are generally not copyrightable content on their
  own, and Nepal does not have an EU-style database right, so the "notspecified" tag
  does not block reuse of the *facts*. The genuinely original content in this dataset
  is the LLM-generated Nepali question/answer phrasing built around those facts.
- This dataset is released under **CC BY 4.0**. Attribution to Nepal's Ministry of
  Education / Open Data Nepal as the source of the underlying statistics is required
  when reusing this dataset.
- This is not legal advice. If you plan to redistribute commercially or at scale,
  independently confirm current Open Data Nepal terms and, if this dataset was
  generated using a third-party LLM API, that provider's usage terms for generated
  output.

## Generation Process

1. Source enrollment figures extracted from a single Ministry of Education CSV
   (district × grade enrollment counts, FY 2011–012 BS), sourced via Open Data Nepal.
2. One numeric fact per row selected as the "grounding" value.
3. An English question/answer pair constructed around that fact.
4. The pair translated and paraphrased into Nepali by an LLM, with the grounded
   numeric value preserved verbatim.
5. Automated audit pass (this document's [Data Quality & Audit](#data-quality--audit)
   section) run against the full output before release.

## Citation

If you use this dataset, please cite it and attribute the underlying statistics to
Nepal's Ministry of Education via Open Data Nepal.

```
@misc{nepali_grounded_education_qa,
  title  = {Nepali Grounded Education QA},
  author = {Tharu, Som Narayan},
  year   = {2026},
  note   = {Derived from Nepal Ministry of Education enrollment statistics, Open Data Nepal},
  url    = {<add your Hugging Face repo URL here>}
}
```