theguywithblacktie/hinglish-conversations

🇮🇳 Hinglish Conversations & Instructions Dataset

A high-purity, multi-subset conversational and instruction-following dataset curated for fine-tuning Large Language Models (such as LLaMA-3 / LLaMA-3.2, Mistral, and Qwen) to converse naturally, fluently, and authentically in Romanized Hinglish (code-mixed Hindi and English in Latin script).

---

📌 Key Highlights

• 100% Verified Hinglish: Every conversation turn and assistant output is strictly filtered to ensure genuine Romanized Hindi vocabulary and cultural expressions. Pure English dialogues and un-translated benchmark remnants have been completely removed.
• Context-Aware Multi-Turn Support: Complex dialogues are decomposed step-by-step, preserving the full conversational history in the input column so the model learns context retention and multi-turn coherence.
• Rich Dialogue Depth: Short, robotic one-liners have been filtered out ($\ge 100$ characters on open-domain conversations), prioritizing expressive, multi-sentence discussions on daily life, tech, work, food, travel, and cultural topics.
• Two Specialized Subsets: Includes a fast-iteration small subset (50k pairs) and a comprehensive large subset (405k pairs).

---

📊 Dataset Subsets & Statistics

Subset	Train Split	Test Split	Total Rows	Target Use Case
small	45,000	5,000	50,000	Fast experimentation, hyperparameter tuning, and rapid loss validation (~35–45 mins on free T4 GPU).
large	364,841	40,538	405,379	Full production fine-tuning across all domains for maximum vocabulary and dialogue depth.

---

📂 Data Format & Schema

Every row in both small and large subsets adheres to a structured, OpenAI / LLaMA-3 compatible list-of-dicts message format:

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

Multi-Turn Example (Context History Preserved)

{
  "input": [
    {
      "role": "user",
      "content": "I sat in a lot of traffic the other day as there was a wreck up ahead. This crazy person decided to zoom by in the emergency lane."
    },
    {
      "role": "assistant",
      "content": "Arre yaar! That's so frustrating! People are so mad, right? The emergency lane is for actual emergencies, not to cut lines. Tu chupke se unka video le leta, then post it online! 😂"
    },
    {
      "role": "user",
      "content": "Ya, it was just unbelievable! I was so annoyed. I wanted to say something, but then thought better of it."
    }
  ],
  "output": [
    {
      "role": "assistant",
      "content": "Haan, it's better to not get involved bhai, especially when people are behaving like that. Safety first! It's good tu didn't say anything. Just let them go, and keep your peace of mind."
    }
  ]
}

---

🚀 Quickstart & Usage

Loading the dataset in Python using Hugging Face datasets:

from datasets import load_dataset

# 1. Load the Small subset (50,000 rows)
small_data = load_dataset("theguywithblacktie/hinglish-conversations", "small")
print(small_data)
# Output: DatasetDict({ train: 45000 rows, test: 5000 rows })

# 2. Load the Large subset (405,379 rows)
large_data = load_dataset("theguywithblacktie/hinglish-conversations", "large")
print(large_data)
# Output: DatasetDict({ train: 364841 rows, test: 40538 rows })

# 3. Accessing individual splits
train_dataset = small_data["train"]
eval_dataset = small_data["test"]

---

🛠️ Sources & Data Curation

This dataset synthesizes and filters multiple high-quality Hinglish corpora:

• Hinglish Saathi Diverse & Multi-turn: Code-mixed multi-turn conversational exchanges filtered for authentic Hinglish phrasing.
• Databricks Dolly 15k (Hinglish Codemix): Complex instruction-following, summarization, and reasoning tasks in Hinglish.
• DSMJ910 & Sujalvc Instruct: Daily scenarios, tasks, and persona-driven Q&A.
• Synthetic & Compiled Conversations: Everyday Indian cultural dialogues, festive planning, culinary debates, and conversational banter.