README.md

---
language:
  - ne
language_code:
  - npi
pretty_name: Nepali Social SFT Dataset
tags:
  - nepali
  - nepal
  - nepali-language
  - devanagari
  - sft
  - supervised-fine-tuning
  - instruction-following
  - social-science
  - synthetic
  - question-answering
  - multiple-choice
task_categories:
  - question-answering
  - text-generation
  - text-classification
license: apache-2.0
size_categories:
  - 10K<n<100K
---
# Nepali Social Studies MCQ — SFT Dataset

A cleaned, deduplicated, bias-corrected instruction-tuning dataset of Nepali-language
multiple-choice questions on social studies topics, derived from the Aya Dataset.

---

## Dataset Summary

| | |
|---|---|
| **Rows** | 27,891 |
| **Language** | Nepali (`ne` / `npi`), Devanagari script |
| **Task type** | Instruction-following (single-turn MCQ Q&A) |
| **Domain** | Social studies (सामाजिक) — MCQ only |
| **License** | Apache-2.0 (permissive) |
| **Source** | `CohereLabs/aya_dataset` (config: `default`, split: `train`), revision `f9ea04583f02a8f86404ff6c58bf75fe637df8a2` |
| **Source subset** | `aya_human_nepali` |
| **Generation type** | Synthetic (originally human-authored source, machine-processed pipeline) |
| **Format** | JSONL, one JSON object per line |

This is a processed derivative of a single upstream slice (`aya_human_nepali`).
It is **not** a general-purpose Nepali instruction dataset — every row is a 4-option
social studies multiple-choice question.

---

## File Structure

Each line is a JSON object:

```json
{
  "id": "sg_25728771e8644154729b5458133abea1",
  "conversations": [
    {"from": "human", "value": "समाजमा शान्ति कायम गर्न के आवश्यक छ? क) द्वन्द्व ख) घृणा ग) ईर्ष्या घ) सद्भाव"},
    {"from": "gpt", "value": "घ) सद्भाव"}
  ],
  "source": "CohereLabs/aya_dataset:default:train",
  "source_name": "aya_human_nepali",
  "source_repo": "CohereLabs/aya_dataset",
  "source_config": "default",
  "source_split": "train",
  "source_revision": "f9ea04583f02a8f86404ff6c58bf75fe637df8a2",
  "source_row_id": "sg_25728771e8644154729b5458133abea1:1",
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "license": "Apache-2.0",
  "license_tier": "permissive",
  "task_type": "instruction-following",
  "generation_type": "synthetic",
  "condition": "synthetic",
  "url": "",
  "metadata_json": "{\"generation_domain\": \"सामाजिक\", \"generation_category\": \"सामाजिक\", \"question_type\": \"बहुविकल्पीय\", \"question_length\": \"अति छोटो तथा छोटो\", \"content_language\": \"नेपाली\", \"content_script\": \"देवनागरी\"}"
}
```

### Field reference

| Field | Type | Notes |
|---|---|---|
| `id` | string | Unique per row. No duplicates. |
| `conversations` | array[2] | Exactly one `human` turn (question + 4 options) and one `gpt` turn (labeled answer). |
| `source*` | string | Full upstream provenance chain, constant across the dataset (single source). |
| `language` / `language_code` / `script` | string | `ne` / `npi` / `Deva` for every row. |
| `license` / `license_tier` | string | `Apache-2.0` / `permissive` for every row. |
| `task_type` | string | `instruction-following` for every row. |
| `generation_type` / `condition` | string | `synthetic` for every row. |
| `url` | string | Always empty — no upstream URL was recorded for this source. |
| `metadata_json` | string (JSON-encoded) | See below. Must be `json.loads`'d — it's stored as a string, not a nested object. |

### `metadata_json` sub-fields (all constant across the dataset)

| Sub-field | Value | Meaning |
|---|---|---|
| `generation_domain` | सामाजिक | Social studies |
| `generation_category` | सामाजिक | Social studies |
| `question_type` | बहुविकल्पीय | Multiple choice |
| `question_length` | अति छोटो तथा छोटो | Very short / short |
| `content_language` | नेपाली | Nepali |
| `content_script` | देवनागरी | Devanagari |

### MCQ format convention

- Options are always labeled `क)` `ख)` `ग)` `घ)` (Devanagari equivalents of A/B/C/D), embedded in the `human` turn after the question stem.
- The `gpt` turn is always `<label>) <option text>` — the label plus the exact option text, space-separated.

---

## Statistics

| Metric | Value |
|---|---|
| Total rows | 27,891 |
| Question length (chars) | min 51 · median 105 · mean 105.5 · max 185 |
| Answer length (chars) | min 5 · median 18 · mean 18.2 · max 53 |
| Unique question strings | 27,889 (2 collisions — see Known Issues) |
| Answer-option label distribution | क) 21.9% · ख) 27.8% · ग) 25.4% · घ) 24.8% |

The answer-label distribution is intentionally near-uniform (see Processing History).

---

## Processing History

This file is the output of a 3-stage cleaning pipeline applied to a raw
`aya_human_nepali` export (originally 29,029 rows):

**Stage 1 — Structural validation**
Verified JSON validity, schema consistency, non-empty turns, and unique IDs.
No rows dropped at this stage.

**Stage 2 — Foreign-script / homoglyph contamination removal**
The raw export had characters from ~18 unrelated Unicode scripts (Armenian,
Gujarati, Gurmukhi, Greek, Arabic, Bengali, Cyrillic, Telugu, Hangul, Kannada,
Malayalam, Hebrew, Sinhala, Georgian, Thai, Ethiopic, Oriya) substituted into
what should have been pure Devanagari text — e.g. `रहित` corrupted to `रहಿತ`.
**1,117 rows** were auto-removed for this reason (a smaller number of visually
similar corruptions remain — see Known Issues).

**Stage 3 — MCQ answer-position rebalancing**
The raw export had 84.8% of correct answers sitting in option क) (position 1)
— a positional bias a model would learn to exploit instead of reading the
question. Each valid, uncorrupted MCQ had its 4 options deterministically
shuffled (seeded by row `id`, reproducible) and the question/answer text
rewritten to match. **27,891 rows** were successfully rebalanced; 21 rows
where the correct option couldn't be confidently identified (corrupted text)
were excluded rather than guessed.

`29,029 → −1,117 (contamination) → −21 (unresolvable MCQ) → 27,891 final rows`

---

## Known Issues (as of this file)

These are documented, not hidden — check before using for training or eval.

1. **Residual character-level corruption (~135 rows / 0.5%)**
   The contamination filter used in Stage 2 didn't cover every Unicode block.
   Confirmed residual cases:
   - Myanmar vowel signs (e.g. `भूमिकသ` — U+1031)
   - Arabic Presentation Forms-B (e.g. a stray U+FEEC inside a Nepali word)
   - Latin Extended-A ligatures/diacritics (`ğ`, `œ`) inside Nepali words
   - One Private Use Area character (U+F8FF)
   - Stray combining diacritics with no base character (U+0308)
   These rows are not flagged in this file and should be filtered before
   training if exact cleanliness matters.

2. **2 duplicate question pairs introduced by rebalancing**
   Two pairs of rows shared the same question stem and same 4-option pool in
   the source data (just in different original order/answer) — a form of
   near-duplication the literal-string dedup step didn't catch pre-rebalance.
   Independent shuffling coincidentally produced identical final text for
   each pair:
   - `sg_870ce75e2773760fbd9b4cb2b69cd555` / `sg_ebe54de3e99d4de5f124c9ca0dd83d14`
   - `sg_e7783f31aec816aef8f9d6786a267c0c` / `sg_8c4838103eb655741dc0fb744d2bf5b5`

3. **Zero diversity by design, not by accident**
   100% of rows are: single source (`aya_human_nepali`), single domain
   (सामाजिक), single question type (MCQ), single length bucket (short).
   This is a narrow, homogeneous slice. If broader Nepali SFT coverage is
   the goal, this file needs to be combined with other domains/sources —
   it is not a general-purpose instruction dataset on its own.

4. **No held-out split**
   All 27,891 rows are from `source_split: train`. There is no dev/test
   split in this file — carve one out before using for evaluation.

---

## Recommended Use

- Suitable as one ingredient in a larger Nepali instruction-tuning mix,
  specifically for MCQ-style social studies knowledge.
- Not suitable on its own for general instruction-following, open-ended
  generation, or any domain outside social studies MCQs — the model will
  overfit to this narrow format if trained on it in isolation.
- Filter or manually review the ~135 residual-corruption rows and the 2
  duplicate pairs above before final training use.

---

## License

Apache-2.0, inherited from the upstream `CohereLabs/aya_dataset`. Verify
this still applies to your specific use case and jurisdiction — Apache-2.0
covers the dataset structure/text; check upstream terms for any additional
conditions CohereLabs may have attached to the Aya Dataset specifically.

---

## Citation / Provenance

If publishing or citing this dataset, credit the upstream source:

```
Source: CohereLabs/aya_dataset (config: default, split: train)
Revision: f9ea04583f02a8f86404ff6c58bf75fe637df8a2
Subset: aya_human_nepali
```

---

## Changelog

| Version | Rows | Change |
|---|---|---|
| Raw export | 29,029 | Original `aya_human_nepali` pull |
| v1 cleaned | 29,029 | Structural validation only (no removals) |
| v2 cleaned | 27,891 | + foreign-script contamination removal (−1,117) + MCQ rebalancing (−21 unresolvable) |