Kashif786/Sindhi-Assistant-IFT-1500

Sindhi Assistant IFT 1,500: Grounded Instruction & Cultural Roleplay Dataset

Sindhi Assistant IFT 1,500 is a production-grade, audited Instruction Fine-Tuning (IFT / SFT) dataset developed specifically for training and fine-tuning long-context multilingual language models in the Sindhi language (سنڌي). It is tailored as the companion alignment dataset for Kashif786/gemma-4-E2B-it-text-only-128k-pruned.

The dataset spans 1,500 master conversations, 22,012 dialogue turns, and 654,305 words, covering 380+ distinct academic, cultural, legal, technical, and colloquial domains.

---

📊 Dataset Summary

Attribute	Value	Description
Total Conversations	1,500	Clean, deduplicated, and structurally verified
Total Dialogue Turns	22,012	Clean role-alternating turns
Total Word Count	654,305	Native Sindhi, English, and Urdu vocabularies
Single-Turn Instructions	806 (53.7%)	Professional AI Q&A with markdown, tables, and code
Multi-Turn Roleplay	694 (46.3%)	Grounded dialogues in Sindhi culture & everyday life
Digit Standard	100% ASCII (0-9)	Zero Arabic-Indic digits ([\u0660-\u0669\u06f0-\u06f9])
Role Alternation	100% Compliant	Every conversation begins with user and ends with assistant
License	Apache 2.0	Permissive open-source license for research and commercial use

---

⚖️ Dual-Subset Architecture: Mitigating Persona Bleeding

A frequent challenge in conversational LLM fine-tuning is persona bleeding—where colloquial roleplay habits (e.g., informal village greetings or poetic emotional reflections) leak into formal technical tasks, or conversely, where robotic structured lists leak into warm social dialogues.

To prevent this, the dataset is cleanly partitioned into two distinct sub-configurations, each paired with its own dedicated system prompt:

1. instruct Subset (806 Conversations — 53.7%)

• Focus: Objective instruction-following, computer science, medicine, Sindh legal acts, economics, mathematics, and bidirectional translations.
• Persona: Authoritative, structured, and formal AI assistant. Responses feature markdown headers, bullet lists, LaTeX/ASCII equations, and clean code blocks.
• System Prompt:

  You are a knowledgeable, respectful, and professional Sindhi AI Assistant.
  

2. cultural_roleplay Subset (694 Conversations — 46.3%)

• Focus: Deep multi-turn dialogues grounded in Sindhi classical poetry (Shah Jo Risalo, Sachal Sarmast, Shaikh Ayaz), agrarian life along the River Indus, folk traditions, social etiquette, and regional history.
• Turn Distribution:
  • Epic Dialogues (50+ turns): 216 conversations
  • Substantial Dialogues (30–49 turns): 149 conversations
  • Natural Multi-turn (3–29 turns): 329 conversations
• Persona: Culturally authentic Sindhi conversational partner participating in natural human dialogue.
• System Prompt:

  You are a culturally grounded Sindhi conversational partner participating in authentic dialogue.
  

---

🔬 Domain Coverage (380+ Unique Knowledge Areas)

The dataset covers six major thematic pillars:

1. Sindhi Classical Literature & Prosody:
   In-depth analysis of all 30 Surs of Shah Abdul Latif Bhittai (Kalyan, Sarang, Sasui, Marui, Kamod, Sorath, etc.), Sachal Sarmast's mystical Kafis, Shaikh Ayaz's modern poetry, and the metrical structures of Sindhi Bait and Wai.
2. Sindh Legal & Administrative Governance:
   Sindh Rented Premises Ordinance 1979, Sindh Public Procurement Regulatory Authority (SPPRA) Rules, Land Revenue Records (Form VII, VIII-A), Sindh Right to Information (RTI) Act 2016, Succession & Legal Heirship procedures, and family court jurisprudence.
3. Computer Science, Software Engineering & AI:
   Python FastAPI asynchronous architectures, PostgreSQL indexing (B-Tree vs GIN), CAP Theorem, Transformer self-attention mechanisms, Docker multi-stage builds, Redis caching patterns, and REST API design.
4. Cross-Lingual & Trilingual Bridges:
   Bidirectional English-to-Sindhi and Sindhi-to-English translations; Urdu-to-Sindhi and Sindhi-to-Urdu cultural exchanges; and comprehensive trilingual technical glossaries (English–Urdu–Sindhi).
5. Medicine, Pharmacology & Public Health:
   Type 2 diabetes management, cardiovascular hemodynamics, chronic kidney disease protocols, snakebite antivenom (ASV) guidelines, thalassemia carrier screening, and dengue therapeutics.
6. Ecology, Agriculture & Indus Hydrology:
   Indus Delta seawater intrusion, Manchar Lake restoration, 1991 Water Apportionment Accord, canal irrigation systems, saline soil remediation, and mangrove conservation.

---

🛠️ Quickstart: Loading the Dataset

Install the Hugging Face datasets library:

pip install datasets

Loading Specific Configurations

from datasets import load_dataset

# 1. Load the unified dataset (all 1,500 conversations):
dataset_all = load_dataset("Kashif786/Sindhi-Assistant-IFT-1500", "all")
print(f"Total records: {len(dataset_all['train'])}")

# 2. Load strictly the 806 AI Instruction-Following tasks:
dataset_instruct = load_dataset("Kashif786/Sindhi-Assistant-IFT-1500", "instruct")
print(f"Instruct records: {len(dataset_instruct['train'])}")

# 3. Load strictly the 694 Cultural Multi-turn Roleplay conversations:
dataset_roleplay = load_dataset("Kashif786/Sindhi-Assistant-IFT-1500", "cultural_roleplay")
print(f"Roleplay records: {len(dataset_roleplay['train'])}")

Example Data Record

sample = dataset_all["train"][0]
print("ID:", sample["id"])
print("Subset:", sample["subset"])
print("Total Turns:", sample["turns_count"])

for turn in sample["messages"]:
    print(f"[{turn['role'].upper()}]: {turn['content'][:80]}...")

---

📂 Repository Structure & Format Compatibility

The repository provides files in both Hugging Face native multi-subset format and industry-standard training formats:

Kashif786/Sindhi-Assistant-IFT-1500/
├── README.md                          # Comprehensive English Dataset Card
├── data/
│   ├── all.jsonl                      # 1,500 unified records (Default Config)
│   ├── instruct.jsonl                 # 806 single-turn instruction QA records
│   └── cultural_roleplay.jsonl        # 694 multi-turn dialogue records
└── formats/
    ├── sindhi_gemma_chatml.jsonl      # Native Gemma ChatML format
    ├── sindhi_sharegpt.jsonl          # ShareGPT format (Axolotl / LLaMA-Factory)
    └── sindhi_alpaca.jsonl            # Alpaca format (Unsloth / standard SFT)

---

🚀 Fine-Tuning Guide (Gemma-4-E2B-it-128k)

Here is a ready-to-run fine-tuning script utilizing Hugging Face transformers and trl:

from datasets import load_dataset
from transformers import AutoTokenizer, AutoModelForCausalLM
from trl import SFTTrainer, SFTConfig

model_id = "Kashif786/gemma-4-E2B-it-text-only-128k-pruned"
tokenizer = AutoTokenizer.from_pretrained(model_id)

# 1. Load the dataset
dataset = load_dataset("Kashif786/Sindhi-Assistant-IFT-1500", "all", split="train")

# 2. Apply Gemma chat template
def format_prompts(example):
    return {"text": tokenizer.apply_chat_template(example["messages"], tokenize=False)}

dataset = dataset.map(format_prompts)

# 3. Configure training arguments
sft_config = SFTConfig(
    output_dir="./sindhi-gemma-sft",
    max_seq_length=4096,
    per_device_train_batch_size=2,
    gradient_accumulation_steps=8,
    learning_rate=2e-5,
    num_train_epochs=3,
    logging_steps=10,
    save_strategy="epoch",
    fp16=True,
    dataset_text_field="text"
)

# 4. Initialize model & trainer
model = AutoModelForCausalLM.from_pretrained(model_id, torch_dtype="auto", device_map="auto")
trainer = SFTTrainer(
    model=model,
    train_dataset=dataset,
    args=sft_config
)

trainer.train()

---

🔍 Quality Assurance & Sanitization Pipeline

1. Strict ASCII Digits Rule:
   All numeric entities across all 654,305 words are standardized to international ASCII digits 0-9. A regex audit for Eastern Arabic-Indic digits ([\u0660-\u0669\u06f0-\u06f9]) returns exactly 0 occurrences.
2. Deterministic Role Alternation:
   Every conversation begins with role: user, strictly alternates between user and assistant, and terminates on role: assistant. Consecutive identical turns were merged and dangling turns resolved.
3. Orthographic Standardization:
   All Sindhi text is normalized to eliminate Unicode character collisions across Perso-Arabic codepoints (standardizing Kaf ڪ, Heh ه, and Yeh ي).

---

📜 Citation

If you use this dataset in your research, evaluations, or model fine-tuning, please cite:

@dataset{kashif2026sindhi_ift,
  author       = {Kashif Ali},
  title        = {Sindhi Assistant IFT 1,500: Grounded Instruction & Cultural Roleplay Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  howpublished = {\url{https://huggingface.co/datasets/Kashif786/Sindhi-Assistant-IFT-1500}}
}

---

👤 Author & Acknowledgments

• Curated and Engineered by: Kashif Ali (@Kashif786)
• Target Model: Kashif786/gemma-4-E2B-it-text-only-128k-pruned
• Dedicated to: The digital preservation, technological empowerment, and high-performance AI representation of the Sindhi language.