sabin1234/Nepali-Science-SFT-Dataset

🇳🇵 Nepali Science SFT Dataset — cleaned_v3

A 50,000-record Nepali Devanagari Supervised Fine-Tuning dataset focused on Science, structured for conversational question–answer training.

---

1. Dataset Information

Dataset identity

Property	Value
Dataset file	cleaned_v3.json
Total records	50,000
Primary domain	Science
Categories	12
Subdomains	79
Language	Nepali (ne)
Script	Devanagari
Primary source model	gemini-3.5-flash-lite
Dataset format	JSON array
Training style	Conversational SFT
Conversation pattern	human → gpt
Unique record IDs	50,000
Duplicate question count	4
Duplicate answer count	132

What this dataset contains

This dataset is designed as a Nepali-language Science SFT corpus. Each record contains a user-style scientific question and a corresponding assistant-style answer. The content spans physics, chemistry, biology, astronomy, genetics, ecology, hydrology, food science, geology, botany, environmental science, and renewable-energy-related topics.

The data is stored as individual conversation examples so it can be adapted to instruction-tuning or conversational supervised fine-tuning workflows.

Language and script

All 50,000 records are tagged as:

• Language: ne
• Script: Devanagari
• Domain: science

The dataset is therefore explicitly labeled as a Nepali Devanagari science corpus at the metadata level.

---

2. Dataset Structure

Each record follows this structure:

{
  "id": "487289ea-414c-47e8-a039-25648a516806",
  "category": "भौतिकशास्त्र",
  "subdomain": "गति र चाल",
  "language": "ne",
  "script": "Devanagari",
  "domain": "science",
  "source_model": "gemini-3.5-flash-lite",
  "conversations": [
    {
      "from": "human",
      "value": "म साइकल चलाउँदा पेडल मार्न छोड्दा पनि साइकल केही समय अगाडि नै गुडिरहन्छ नि, यो गति र चालको कुन नियम वा अवधारणासँग सम्बन्धित छ?"
    },
    {
      "from": "gpt",
      "value": "यो जड़त्वको नियमसँग सम्बन्धित छ। यसलाई चालमा रहेको वस्तुको जड़त्व भनिन्छ..."
    }
  ]
}

Field-by-field explanation

Field	Type	Description
id	String	Unique identifier for each dataset record
category	String	Broad science category
subdomain	String	More specific topic within the category
language	String	ISO-style language code; ne = Nepali
script	String	Writing system used by the dataset
domain	String	High-level domain; science
source_model	String	Model associated with the generated record
conversations	Array	Two-message SFT conversation
conversations[].from	String	Conversation role: human or gpt
conversations[].value	String	Actual Nepali question or answer

---

3. Dataset Hierarchy

The dataset uses the following hierarchy:

Science
│
├── Category
│   │
│   ├── Subdomain
│   │   ├── Question
│   │   └── Answer
│   │
│   └── ...
│
└── ...

Taxonomy totals

1 Domain
   └── Science

12 Categories
   └── 79 Unique Subdomain Labels

50,000 Conversation Records
   ├── Human Question
   └── GPT Answer

---

4. Domain Coverage

Primary domain: Science

Domain	Records	Share
Science	50,000	100.00%

The current file is therefore a single-domain dataset, not a mixed-domain corpus.

---

5. Category Distribution

The dataset contains 12 categories.

#	Category	Records	Share
1	भौतिकशास्त्र	20,042	40.08%
2	रसायनशास्त्र	13,200	26.40%
3	जीवविज्ञान	11,980	23.96%
4	खगोलशास्त्र	4,362	8.72%
5	आनुवंशिकी	302	0.60%
6	पारिस्थितिकी	22	0.04%
7	खाद्य विज्ञान	20	0.04%
8	भूविज्ञान	20	0.04%
9	वनस्पति विज्ञान	20	0.04%
10	जलविज्ञान	12	0.02%
11	पर्यावरण विज्ञान	10	0.02%
12	नवीकरणीय ऊर्जा विज्ञान	10	0.02%
	Total	50,000	100.00%

Category-level interpretation

• Physics is the largest section with 20,042 records.
• Chemistry contributes 13,200 records.
• Biology contributes 11,980 records.
• Astronomy contributes 4,362 records.
• The remaining categories provide smaller specialist topic collections.

Together, Physics, Chemistry, Biology, and Astronomy account for 49,584 records, or 99.17% of the dataset.

---

6. Complete Subdomain Coverage

The dataset contains 79 distinct subdomain labels across 80 category–subdomain assignments. The difference occurs because प्रकाश संश्लेषण appears under both Biology and Botany.

Physics — 21 subdomains

Subdomain	Records
अर्धचालक	1,000
अपवर्तन	1,000
काम र ऊर्जा	1,000
गति र चाल	1,000
गुरुत्वाकर्षण	1,000
घनत्व	1,000
चुम्बकत्व	1,000
दाब	1,015
ताप र तापक्रम	1,000
ताप तथा तापक्रम	1
तरंग	1,000
ध्वनि	1,000
बल र गति	1,000
प्रकाश	1,008
परावर्तन	1,000
लेन्स	1,000
विद्युत	1,000
विद्युतचुम्बकत्व	1,000
विद्युत परिपथ	1,018
शक्ति	1,000
ऊष्मा स्थानान्तरण	1,000
Physics total	20,042

Note: ताप र तापक्रम and ताप तथा तापक्रम are separate metadata values in the current JSON, with the latter containing one record.

Chemistry — 20 subdomains

Subdomain	Records
अकार्बनिक रसायन	600
अक्सिडेसन र रिडक्सन	600
अम्ल र क्षार	600
अणु र यौगिक	1,000
कार्बनिक रसायन	600
ग्यास	600
जैवरसायन	600
द्रावण	600
धातु र अधातु	600
तत्व र आवर्त सारणी	1,000
पेट्रोकेमिकल	600
बहुलक	600
परमाणु संरचना	1,000
रासायनिक गतिविज्ञान	600
रासायनिक प्रतिक्रिया	600
रासायनिक बन्धन	600
रासायनिक सन्तुलन	600
लवण	600
विद्युत रसायन	600
तापरसायन	600
Chemistry total	13,200

Biology — 20 subdomains

Subdomain	Records
अन्तःस्रावी प्रणाली	600
आनुवंशिकी	600
कोशिका	600
कोशिका विभाजन	600
डीएनए र आरएनए	600
पाचन प्रणाली	600
प्रजनन प्रणाली	600
प्रतिरक्षा प्रणाली	600
प्रोटिन संश्लेषण	600
ब्याक्टेरिया	600
भाइरस	600
मानव शरीर	600
रक्तसञ्चार प्रणाली	600
वनस्पति प्रजनन	600
वनस्पति संरचना	600
विकासवाद	600
श्वसन प्रणाली	600
सूक्ष्मजीव	600
स्नायु प्रणाली	600
प्रकाश संश्लेषण	580
Biology total	11,980

Astronomy — 11 subdomains

Subdomain	Records
आकाशगंगा	200
क्षुद्रग्रह	459
ग्रह	600
चन्द्रमा	600
तारा	200
तारामण्डल	200
धूमकेतु	200
सूर्य	600
सौर्यमण्डल	600
उपग्रह	600
ब्ल्याक होल	103
Astronomy total	4,362

Genetics — 1 subdomain

Subdomain	Records
डीएनए	302
Genetics total	302

Other specialist categories

Category	Subdomain	Records
पारिस्थितिकी	खाद्य शृङ्खला	22
खाद्य विज्ञान	खाद्य संरक्षण	20
भूविज्ञान	प्लेट टेक्टोनिक्स	20
वनस्पति विज्ञान	प्रकाश संश्लेषण	20
जलविज्ञान	जलचक्र	12
पर्यावरण विज्ञान	जलवायु परिवर्तन	10
नवीकरणीय ऊर्जा विज्ञान	जलविद्युत	10

---

7. Conversation Format

Every record contains exactly two conversation messages:

1. human
2. gpt

This produces the basic SFT pattern:

Instruction / Question
        ↓
Assistant Response

Structural audit

Check	Result
Records	50,000
Records with 2-message conversations	50,000
Expected role order	human → gpt
Records with invalid conversation structure	0
Records with missing top-level fields	0
Unique IDs	50,000
Duplicate IDs	0

---

8. Text Characteristics

The final file contains:

• Nepali-language questions
• Nepali-language explanatory answers
• Devanagari-script metadata
• Natural-language questions rather than only short factual prompts
• Explanatory answers with examples and reasoning
• Scientific terminology across multiple disciplines
• Numerical and formula-based explanations in relevant science topics

Question length

Statistic	Characters
Minimum	15
Maximum	264
Average	65.50
Median	63

Answer length

Statistic	Characters
Minimum	30
Maximum	1,051
Average	206.81
Median	194

This gives the corpus a relatively broad answer-length distribution rather than forcing every answer into the same short template.

---

9. Cleaning and Normalization Coverage

The cleaned_v3.json artifact was checked at both the schema level and the text level.

Structural cleaning / validation

The final file passed the following structural checks:

• Valid JSON parsing
• Exactly 50,000 records
• Consistent top-level schema across all records
• Unique record IDs
• No missing IDs
• Consistent conversation structure
• Exactly two messages per conversation
• Correct human → gpt role ordering
• Non-empty question and answer values
• Consistent metadata fields
• Consistent language tag
• Consistent script tag
• Consistent domain tag

Unicode / text normalization

The final artifact was checked for:

• Unicode NFC normalization
• Control characters
• Basic whitespace anomalies
• Text encoding integrity

Current audit result:

Text-quality check	Result
NFC-normalization issues	0
Disallowed control-character rows	0
Whitespace-suspect rows	4
Questions ending with ?	50,000 / 50,000

---

10. Question-Marking Validation

A dedicated question-mark check was applied to the human side of every conversation.

Result

50,000 / 50,000 questions end with ?.

That means the current dataset has complete question-mark coverage on the question field.

This is important because a conversational SFT dataset should clearly distinguish the user query from the answer.

---

11. Duplicate Detection

Duplicate checks were performed separately for:

• Question text
• Answer text
• Record IDs

Results

Duplicate check	Count
Duplicate IDs	0
Duplicate questions	4
Duplicate answers	132

The very low number of duplicate questions compared with the total corpus indicates that the question set is largely distinct, while repeated answers occur more often because scientifically similar questions can naturally map to similar explanatory responses.

---

12. Language Character Audit

The dataset is labeled as Nepali + Devanagari in its metadata, but a raw-character audit of the conversation text found 2,265 records containing at least one Latin alphabet character.

This means the metadata is fully consistent:

language = ne
script   = Devanagari

but the conversation text itself is not literally restricted to Devanagari-only characters in every row.

The detected Latin characters occur inside scientifically formatted content, abbreviations, formulas, symbolic notation, or model-generated text.

Therefore, the most accurate description of the current artifact is:

Nepali Devanagari-focused Science SFT dataset with some Latin/symbolic scientific notation remaining inside conversation text.

This distinction is important for reproducible dataset documentation.

---

13. Answer Punctuation Audit

The answer field was also checked for sentence-ending punctuation.

• 49,930 answers end with ।
• 70 answers end with another character or punctuation pattern.

The non-। endings include cases involving mathematical notation, brackets, symbols, formulas, or other generated formatting.

This does not affect JSON validity or conversation structure, but it is recorded here so the README reflects the actual state of the final file.

---

14. Source Model Distribution

The source_model metadata is not completely uniform.

Source model	Records	Share
gemini-3.5-flash-lite	49,512	99.02%
gemini-3.6-flash	487	0.97%
himalaya-gemma	1	<0.01%
Total	50,000	100.00%

Therefore, although gemini-3.5-flash-lite is the dominant generator, the dataset is not technically a single-source-model corpus.

---

15. Scientific Coverage

The dataset covers a broad range of scientific concepts.

Physics coverage

The Physics section includes:

• Motion and speed
• Force and motion
• Work and energy
• Power
• Gravitation
• Density
• Pressure
• Heat and temperature
• Heat transfer
• Sound
• Waves
• Light
• Reflection
• Refraction
• Lenses
• Electricity
• Electrical circuits
• Magnetism
• Electromagnetism
• Semiconductors

Chemistry coverage

The Chemistry section includes:

• Atomic structure
• Elements and periodic table
• Molecules and compounds
• Chemical bonding
• Chemical reactions
• Acids and bases
• Salts
• Oxidation and reduction
• Organic chemistry
• Inorganic chemistry
• Solutions
• Gases
• Metals and non-metals
• Chemical equilibrium
• Chemical kinetics
• Thermochemistry
• Electrochemistry
• Biochemistry
• Petrochemicals
• Polymers

Biology coverage

The Biology section includes:

• Cell biology
• Cell division
• Genetics
• DNA and RNA
• Protein synthesis
• Human body
• Digestive system
• Respiratory system
• Blood circulation
• Nervous system
• Endocrine system
• Reproductive system
• Immune system
• Viruses
• Bacteria
• Microorganisms
• Evolution
• Plant structure
• Plant reproduction
• Photosynthesis

Astronomy coverage

The Astronomy section includes:

• Sun
• Moon
• Planets
• Solar system
• Satellites
• Stars
• Constellations
• Galaxies
• Asteroids
• Comets
• Black holes

---

16. Data Quality Summary

Final audit snapshot

Dimension	Status
JSON structure	✅ Valid
Record count	✅ 50,000
Unique IDs	✅ 50,000
Conversation structure	✅ Consistent
Human/GPT ordering	✅ Consistent
Empty questions	✅ None detected
Empty answers	✅ None detected
Question marks	✅ 50,000 / 50,000
NFC normalization	✅ No issues detected
Control characters	✅ None detected
Duplicate questions	⚠️ 4
Duplicate answers	⚠️ 132
Latin characters in text	⚠️ 2,265 rows
Whitespace-suspect rows	⚠️ 4
Non-standard answer endings	⚠️ 70
Language metadata	✅ ne
Script metadata	✅ Devanagari
Domain metadata	✅ science

---

17. Dataset Statistics at a Glance

┌──────────────────────────────────────────────────────┐
│              NEPALI SCIENCE SFT DATASET              │
├──────────────────────────────────────────────────────┤
│ Records                         50,000                │
│ Domain                               1               │
│ Categories                          12               │
│ Subdomains                          79               │
│ Language                            ne               │
│ Script                       Devanagari              │
│ Unique IDs                      50,000               │
│ Conversation format       human → gpt               │
│ Question duplicates                  4               │
│ Answer duplicates                  132               │
└──────────────────────────────────────────────────────┘

---

18. Recommended Dataset Loading Pattern

A basic Python loading example:

import json

with open("cleaned_v3.json", "r", encoding="utf-8") as f:
    dataset = json.load(f)

print("Total records:", len(dataset))

sample = dataset[0]

print(sample["category"])
print(sample["subdomain"])
print(sample["conversations"][0]["value"])
print(sample["conversations"][1]["value"])

---

19. SFT Training Representation

For a conversational SFT pipeline, one example can be interpreted as:

USER:
<value from conversations[0]>

ASSISTANT:
<value from conversations[1]>

The metadata fields can be retained for:

• Dataset filtering
• Topic balancing
• Evaluation splits
• Subdomain-specific experiments
• Error analysis
• Sampling
• Curriculum design
• Model comparison

---

20. Dataset Design Summary

The current dataset has a clear separation between:

Metadata layer

id
category
subdomain
language
script
domain
source_model

Training-content layer

conversations
    ├── human
    └── gpt

This design makes the dataset useful both as a training corpus and as a structured research dataset where individual records can be filtered by scientific topic.

---

21. File Properties

Property	Value
Filename	cleaned_v3.json
File format	JSON
Encoding	UTF-8
Records	50,000
Approximate file size	55.6 MB
Top-level object	JSON array
Record object	JSON object
Conversation entries	2 per record

---

22. Example Record

{
  "id": "c3ae541f-e257-4d19-ad06-602dd93053a5",
  "category": "भौतिकशास्त्र",
  "subdomain": "गति र चाल",
  "language": "ne",
  "script": "Devanagari",
  "domain": "science",
  "source_model": "gemini-3.5-flash-lite",
  "conversations": [
    {
      "from": "human",
      "value": "गाडीमा गति माप्ने यन्त्रलाई स्पिडोमिटर भनिन्छ भने यसले गाडीको कुन प्रकारको चाल देखाउँछ?"
    },
    {
      "from": "gpt",
      "value": "स्पिडोमिटरले गाडीको तात्कालिक चाल देखाउँछ। यसको अर्थ गाडी कुन निश्चित क्षणमा कति गतिमा चलिरहेको छ भन्ने जानकारी यसले प्रदान गर्छ।"
    }
  ]
}

---

23. Dataset Identity Statement

cleaned_v3.json is a 50,000-record Nepali Science conversational SFT dataset containing 1 science domain, 12 science categories, and 79 subdomains. The dataset uses Nepali (ne) and Devanagari metadata, with a two-turn human → gpt conversation structure and unique record IDs.

The dataset contains substantial coverage of Physics, Chemistry, Biology, and Astronomy, alongside smaller specialist collections in Genetics, Ecology, Food Science, Geology, Botany, Hydrology, Environmental Science, and Renewable Energy Science.

---

24. License / Attribution

This README documents the structure and measured properties of the supplied cleaned_v3.json artifact.

The dataset's actual licensing, redistribution rights, and source-content permissions should be determined from the project's applicable data-generation and source-model terms rather than inferred from the JSON structure itself.

---

Dataset Snapshot

50,000 Records · 1 Domain · 12 Categories · 79 Subdomains · Nepali · Devanagari · Conversational SFT