nassimjp/Bilingual-SFT-2.0-Pashto-English-Aligned

Bilingual SFT 2.0 — Pashto English Aligned 🇦🇫🇬🇧

Bilingual-SFT-2.0-Pashto-English-Aligned is a bilingual supervised fine-tuning dataset designed to improve Large Language Models (LLMs) in Pashto ↔ English understanding, instruction following, conversation, and bilingual generation.

The dataset uses a conversational messages format and is intended for modern instruction-tuning pipelines, including Hugging Face Transformers, TRL, Unsloth, Axolotl, and other SFT frameworks.

🎯 Purpose

The main goal of this dataset is to help language models:

• —Understand Pashto instructions and questions
• —Generate natural Pashto responses
• —Understand English instructions and questions
• —Generate English responses
• —Handle Pashto ↔ English bilingual interactions
• —Improve instruction-following behavior
• —Improve conversational alignment
• —Support Pashto as a low-resource language in modern LLM training

The dataset is particularly useful for models being adapted from primarily English or multilingual foundations toward stronger Pashto-English bilingual capability.

🌍 Languages

The dataset is designed for bilingual training rather than English-only or Pashto-only instruction tuning.

📦 Dataset Format

Each example follows a standard conversational structure:

{
  "messages": [
    {
      "role": "user",
      "content": "پښتو ژبه څه ډول ژبه ده؟"
    },
    {
      "role": "assistant",
      "content": "پښتو د هندو اروپايي ژبو د کورنۍ یوه ژبه ده او د ایراني ژبو له ډلې څخه شمېرل کېږي."
    }
  ]
}

English examples can use the same structure:

{
  "messages": [
    {
      "role": "user",
      "content": "What is Pashto?"
    },
    {
      "role": "assistant",
      "content": "Pashto is an Indo-European language belonging to the Iranian branch."
    }
  ]
}

Bilingual examples may also contain interactions where the user and assistant use different languages.

🧱 Dataset Structure

The primary field is:

The format is compatible with common chat-based SFT pipelines.

🧠 Intended Uses

This dataset can be used for:

Pashto LLM Fine-Tuning

• —Pashto instruction tuning
• —Pashto conversational training
• —Pashto-English bilingual models
• —Pashto language adaptation
• —Low-resource language alignment

English ↔ Pashto Tasks

• —Bilingual question answering
• —Translation-oriented instruction following
• —Cross-lingual understanding
• —Bilingual dialogue
• —Multilingual reasoning

AI Assistants

The conversational structure makes the dataset suitable for training:

• —Chatbots
• —Personal assistants
• —Educational assistants
• —Pashto AI assistants
• —Bilingual customer-support systems
• —Local/offline language models

🚀 Loading the Dataset

Load the dataset with Hugging Face datasets:

from datasets import load_dataset

dataset = load_dataset(
    "nassimjp/Bilingual-SFT-2.0-Pashto-English-Aligned"
)

For the training split:

dataset = load_dataset(
    "nassimjp/Bilingual-SFT-2.0-Pashto-English-Aligned",
    split="train"
)

🔧 Example SFT Preparation

The messages structure can be passed through the tokenizer's chat template:

def format_chat(example):
    return {
        "text": tokenizer.apply_chat_template(
            example["messages"],
            tokenize=False,
            add_generation_prompt=False
        )
    }

dataset = dataset.map(format_chat)

This makes the dataset suitable for many modern SFT training pipelines.

⚡ Unsloth

The dataset can be used with Unsloth-based supervised fine-tuning workflows.

Example:

from datasets import load_dataset

dataset = load_dataset(
    "nassimjp/Bilingual-SFT-2.0-Pashto-English-Aligned",
    split="train"
)

The messages format can then be used with the appropriate chat template for the base model.

🤖 Recommended Training

This dataset is intended for Supervised Fine-Tuning (SFT) rather than continual pretraining.

Recommended workflow:

Base LLM
   ↓
Bilingual SFT 2.0
   ↓
Pashto-English Instruction Tuning
   ↓
Evaluation
   ↓
Optional domain-specific adapters

For best results, combine this dataset with high-quality Pashto conversational, instructional, mathematical, technical, literary, and domain-specific datasets.

🧹 Data Quality

The dataset is intended to provide structured, model-ready conversational examples.

Recommended quality practices when extending or rebuilding the dataset include:

• —Unicode normalization
• —Pashto character preservation
• —Duplicate removal
• —Empty-example removal
• —Language verification
• —Conversation validation
• —Instruction/response consistency checks
• —English leakage detection
• —Pashto orthography checks
• —Length filtering
• —Repeated-example detection

Special attention should be given to Pashto-specific characters such as:

ټ ډ ړ ږ ښ څ ځ ڼ ګ

These characters are important for accurate Pashto representation and tokenizer coverage.

📚 Training Applications

This dataset can be used as one component of a broader Pashto AI training ecosystem.

Possible applications include:

• —Pashto foundation-model adaptation
• —Pashto SFT
• —Bilingual SFT
• —Chat alignment
• —Instruction following
• —Cross-lingual transfer
• —Pashto tokenizer evaluation
• —Pashto-English model evaluation
• —Local and offline AI assistants

⚠️ Limitations

As with any automatically collected or transformed instruction dataset, users should independently evaluate:

• —Translation accuracy
• —Linguistic quality
• —Cultural appropriateness
• —Factual correctness
• —Duplicate or near-duplicate examples
• —Biases inherited from source datasets
• —Distribution between Pashto and English

The dataset should not be treated as a replacement for native-speaker evaluation.

For production systems, additional human review and task-specific evaluation are recommended.

📜 License

This dataset is released under the Apache License 2.0.

Users should also verify the licensing and attribution requirements of any external datasets or source material incorporated into future versions of the dataset.

📖 Citation

If you use this dataset in research or model training, please cite:

@dataset{bilingual_sft_2_pashto_english_2026,
  author       = {Nasibullah Nassim},
  title        = {Bilingual SFT 2.0 — Pashto English Aligned},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/nassimjp/Bilingual-SFT-2.0-Pashto-English-Aligned}
}

👤 Author

Nasibullah Nassim (Nassim JP)

Pashto AI Research & Development

Hugging Face: nassimjp

Project: iPashto.ai

🤝 Contributions

Contributions that improve Pashto-English bilingual AI are welcome.

Useful contributions include:

• —High-quality native Pashto examples
• —Better bilingual alignment
• —Duplicate detection
• —Translation verification
• —Pashto linguistic corrections
• —Dataset quality evaluation
• —Additional instruction categories
• —Better bilingual benchmarks

🌐 Related Pashto AI Resources

This dataset is part of a broader effort to build open Pashto NLP resources, including conversational datasets, instruction datasets, parallel corpora, and continual-pretraining corpora.

Examples of related resources include:

• —nassimjp/pashto-dialogue
• —nassimjp/instruct_pashto
• —nassimjp/english_pashto_parallel-corpus
• —nassimjp/pashto-cpt-corpus
• —nassimjp/pashto_open-perfectblend

The broader objective is to make high-quality Pashto data available for modern open-source AI models.

⭐ Summary

Bilingual SFT 2.0 — Pashto English Aligned provides a conversational foundation for adapting language models toward stronger Pashto-English bilingual instruction following and dialogue.

It is designed to be:

• —🇦🇫 Pashto-aware
• —🇬🇧 English-compatible
• —💬 Conversation-ready
• —🧠 SFT-ready
• —🤖 LLM-friendly
• —🔧 Compatible with modern training frameworks
• —🌍 Useful for low-resource multilingual AI research

Pashto + English → one bilingual training foundation.