sabin1234/nepali-environment-sft

🌿 Nepali Environment Awareness SFT Dataset

A Supervised Fine-Tuning (SFT) dataset for Nepali-language instruction-following, focused entirely on environment awareness topics. All conversations are in Devanagari script, synthetically generated using a Nepali language model, and cleaned for training-ready quality.

---

📋 Dataset Summary

Field	Value
Dataset Name	nepali_environment_awareness_sft
Version	v1.0 (cleaned: v2)
Language	Nepali (ne)
Script	Devanagari (Deva)
Domain	Environment Awareness
Task Type	Instruction Following
Generation Type	Synthetic (Model-Generated)
Total Records	1,000
Train Split	1,000 (100%)
File Size	~3.59 MB
Total Characters	1,138,645
Approx. Tokens	~284,661
License	Apache-2.0 (Permissive)
Format	JSONL (ShareGPT-style conversations)

---

📁 Files

datasets/
└── processed/
    ├── clean_environment_sft.jsonl        # Original (pre-cleaning)
    ├── clean_environment_sft.csv          # Original CSV format
    ├── clean_environment_sft.parquet      # Original Parquet format
    └── clean_environment_sft_v2.jsonl     # ✅ Final cleaned version (use this)

---

🗂️ Data Format

The dataset uses ShareGPT / conversational format — same structure used by platforms like Hugging Face for SFT training with frameworks such as LLaMA-Factory, Axolotl, and TRL.

Each record is a single JSON object per line (.jsonl):

{
  "id": "environment_sft_000001",
  "conversations": [
    {
      "from": "human",
      "value": "ध्वनि प्रदूषण को महत्त्वबारे सरल रूपमा व्याख्या गर्नुहोस्।"
    },
    {
      "from": "gpt",
      "value": "ध्वनि प्रदूषन भनेको अत्यधिक चर्को आवाज वा ध्वनि प्रदूषण हो..."
    }
  ],
  "source": "synthetic",
  "source_name": "nepali_environment_awareness_sft",
  "source_repo": "local_generation",
  "source_config": "environment_awareness",
  "source_split": "train",
  "source_revision": "v1.0",
  "source_row_id": 1,
  "language": "ne",
  "language_code": "npi",
  "script": "Deva",
  "license": "Apache-2.0",
  "license_tier": "permissive",
  "task_type": "instruction-following",
  "generation_type": "synthetic",
  "condition": "model-generated",
  "url": "",
  "metadata_json": "{\"domain\": \"environment_awareness\", \"model\": \"himalaya-gemma-4-e2b-it\", \"generation_number\": 1}"
}

Conversation Roles

Role	Description
human	The instruction / question (Nepali)
gpt	The response / answer (Nepali)

---

📊 Dataset Statistics

Text Length (in characters)

Metric	Human Turn	GPT Turn
Minimum	28	498
Maximum	75	1,705
Average	50.8	1,087.9

Metadata Distribution

Field	Value	Count
Language	ne (Nepali)	1,000
Script	Deva (Devanagari)	1,000
Source	synthetic	1,000
Task Type	instruction-following	1,000
Generation Type	synthetic	1,000
Condition	model-generated	1,000
Domain	environment_awareness	1,000
License	Apache-2.0	1,000
Generator Model	himalaya-gemma-4-e2b-it	1,000

---

🌱 Topics Covered

The dataset covers a wide range of environmental topics including (but not limited to):

• वायु प्रदूषण — Air pollution causes, effects, and prevention
• ध्वनि प्रदूषण — Noise pollution awareness
• जल संरक्षण — Water conservation and sustainable usage
• वन संरक्षण — Forest and jungle conservation
• माटो प्रदूषण — Soil pollution and remediation
• जैविक विविधता — Biodiversity protection
• वन्यजन्तु संरक्षण — Wildlife conservation habits
• सौर ऊर्जा — Solar energy promotion
• नवीकरणीय ऊर्जा — Renewable energy sources
• जलवायु परिवर्तन — Climate change adaptation and mitigation
• कार्बन उत्सर्जन — Carbon emission reduction
• फोहोर व्यवस्थापन — Waste management and recycling
• पुनःप्रयोग — Reuse and reduce practices
• हरित सहर — Green city and sustainable urban development
• नदी संरक्षण — River and water body conservation
• खडेरी — Drought awareness and water management
• प्राकृतिक सम्पदा — Natural resource management

---

🔧 Generation Details

Field	Value
Generator Model	himalaya-gemma-4-e2b-it
Generation Method	Synthetic instruction-response pairs
Source Repo	local_generation
Config	environment_awareness
Revision	v1.0
ID Range	environment_sft_000001 → environment_sft_001000

---

🧹 Data Cleaning Report (v2)

The original file (clean_environment_sft.jsonl) was inspected and cleaned to produce clean_environment_sft_v2.jsonl.

Issues Found & Fixed

Issue	Records Affected	Turns Fixed
Non-Devanagari characters (Malayalam, Bengali, Arabic, Cyrillic, Gujarati, Thai, Korean, Chinese, etc.)	57 records	167 turns
Stray curly braces used as section labels e.g. {वृक्षारोपण: content}	28 records	28 turns
Stray (2) artifacts from CO₂ subscript encoding failure	31 turns	31 turns
Excessive/multiple whitespace	—	151 instances

Non-Devanagari Characters Removed

The following foreign script characters were found and removed (replaced with space to avoid word merging):

Script	Example Characters	Unicode Range
Arabic	ن, ش, و, ر	U+0600–U+06FF
Bengali	ণ, া, ল, ি	U+0980–U+09FF
Malayalam	്, ന, ത, ര	U+0D00–U+0D7F
Gujarati	ત, ગ, ા	U+0A80–U+0AFF
Punjabi (Gurmukhi)	ਣ, ਪ, ਲ	U+0A00–U+0A7F
Telugu	ు, ి	U+0C00–U+0C7F
Cyrillic	и, с	U+0400–U+04FF
Thai	ฎ, ท, ำ	U+0E00–U+0E7F
Chinese	这	U+4E00–U+9FFF
Korean	활	U+AC00–U+D7FF
Japanese	。	U+3000–U+303F

Stray (2) Artifact

The subscript ₂ in CO₂ was corrupted during generation as a literal (2) token. Example:

• Before: कार्बनडाइअक्साइड (2) जस्ता हरितगृह ग्यासहरू
• After: कार्बनडाइअक्साइड जस्ता हरितगृह ग्यासहरू

Since the surrounding Nepali text already names the compound, the (2) was safely removed without loss of meaning.

Stray Brace Fix

Section headers wrapped in curly braces were unwrapped:

• Before: {वृक्षारोपण: वनजङ्गाहरू कार्बनडाइअक्सिड सोसेर...}
• After: वृक्षारोपण: वनजङ्गाहरू कार्बनडाइअक्सिड सोसेर...

Post-Cleaning Validation

Check	Result
Total records	✅ 1,000
Duplicate IDs	✅ 0
Empty human turns	✅ 0
Empty GPT turns	✅ 0
Non-Devanagari characters remaining	✅ 0
Stray braces remaining	✅ 0
Stray (2) remaining	✅ 0

---

🚀 Usage

Load with Python

import json

dataset = []
with open("datasets/processed/clean_environment_sft_v2.jsonl", encoding="utf-8") as f:
    for line in f:
        dataset.append(json.loads(line.strip()))

# Access a record
record = dataset[0]
print(record["conversations"][0]["value"])  # Human question
print(record["conversations"][1]["value"])  # GPT answer

Load with Hugging Face datasets

from datasets import load_dataset

dataset = load_dataset("json", data_files="datasets/processed/clean_environment_sft_v2.jsonl")
print(dataset["train"][0])

Use with LLaMA-Factory / Axolotl (ShareGPT format)

The dataset is in ShareGPT format (conversations list with from/value keys). Register it in your training config:

# dataset_info.json entry
"nepali_environment_sft": {
  "file_name": "clean_environment_sft_v2.jsonl",
  "formatting": "sharegpt",
  "columns": {
    "messages": "conversations"
  }
}

---

📜 License

This dataset is released under the Apache-2.0 license (permissive).

You are free to:

• Use commercially
• Modify and distribute
• Use for model training and fine-tuning

See Apache-2.0 License for full terms.

---

🏷️ Citation

If you use this dataset, please cite:

@dataset{nepali_environment_awareness_sft_2026,
  title        = {Nepali Environment Awareness SFT Dataset},
  language     = {Nepali},
  script       = {Devanagari},
  domain       = {Environment Awareness},
  records      = {1000},
  format       = {ShareGPT / JSONL},
  license      = {Apache-2.0},
  version      = {v1.0},
  generator    = {himalaya-gemma-4-e2b-it},
  year         = {2026}
}

---

📬 Contact

Generated and curated locally. For questions or contributions, refer to the source repository configuration: local_generation / environment_awareness.