📄 goal-pashto-chat-sharegpt-5GB — Pashto ShareGPT‑Style Chat Dataset

A large‑scale, high‑quality Pashto conversational dataset designed for instruction‑tuning, dialogue modeling, and LLM alignment.
This dataset contains ~5GB of multi‑turn Pashto conversations inspired by ShareGPT, covering reasoning, advice, education, culture, and general knowledge.

---

📌 Dataset Summary

goal-pashto-chat-sharegpt-5GB is a curated collection of Pashto user–assistant conversations following the standard chat schema:

{
  "messages": [
    {"role": "user", "content": "..."},
    {"role": "assistant", "content": "..."}
  ]
}

The dataset is optimized for:

• Pashto chatbots
• Instruction‑following LLMs
• Multi‑turn dialogue agents
• Cultural & linguistic alignment
• Long‑context reasoning models (4k–16k)

---

📂 Data Structure

Each entry contains:

Field	Description
messages	List of chat turns
role	"user" or "assistant"
content	Pashto text of the message

Example

{
  "messages": [
    {
      "role": "user",
      "content": "په افغانستان کې د ژمي موسم څنګه وي؟"
    },
    {
      "role": "assistant",
      "content": "په ډېرو سیمو کې ژمی یخ او واورین وي، په ځانګړي ډول په غرنیو ولایتونو کې."
    }
  ]
}

---

📊 Token Length Distribution

This dataset contains short, medium, long, and very long conversations:

Bucket	Count
≤512	45
≤1024	341
≤2048	425
≤4096	340
≤8192	included
≤16384	included
>16384	included

This makes the dataset ideal for training 8k–16k context Pashto LLMs.

---

🧠 Use Cases

• Pashto conversational AI
• Chatbot fine‑tuning (SFT)
• Cultural alignment for LLMs
• Long‑context reasoning
• Multi‑turn dialogue generation
• Low‑resource NLP research

---

🛠 Training Format

Compatible with:

• LLaMA‑3 / LLaMA‑2 chat templates
• Qwen chat template
• Unsloth SFT pipelines
• Talanda data‑engineering framework

Example training snippet

tokenizer.apply_chat_template(
    example["messages"],
    tokenize=True,
    add_generation_prompt=False
)

---

📥 Loading the Dataset

from datasets import load_dataset
ds = load_dataset("nassimjp/goal-pashto-chat-sharegpt-5GB")

---

📚 Citation

@dataset{nassimjp2026goalpashtochat,
  title        = {Goal Pashto Chat ShareGPT 5GB},
  author       = {Nassimjp},
  year         = {2026},
  publisher    = {Hugging Face Datasets},
  url          = {[https://huggingface.co/datasets/nassimjp/goal-pashto-chat-sharegpt-5GB](https://huggingface.co/datasets/nassimjp/goal-pashto-chat-sharegpt-5GB)},
  note         = {A large-scale Pashto conversational dataset for instruction tuning and chat-based LLM training.}
}

---

📜 License

This dataset is released under a non‑commercial research license.

Please review the license terms before using it in production systems.

---

🙌 Acknowledgements

This dataset is part of the broader effort to build high‑quality Pashto NLP resources and support low‑resource language AI development.