qrk-labs/akeel-thought-injection-50k-augmented

<div align="center">

<!-- Muqarnas Layers Logo --> <svg width="80" height="80" viewBox="0 0 64 64" fill="none" xmlns="http://www.w3.org/2000/svg"> <path d="M32 4 L4 32 L32 60 L60 32 Z" fill="none" stroke="#fafafa" stroke-width="0.6" opacity="0.3"/> <path d="M32 10 L10 32 L32 54 L54 32 Z" fill="none" stroke="#fafafa" stroke-width="0.8" opacity="0.5"/> <path d="M32 16 L16 32 L32 48 L48 32 Z" fill="none" stroke="#fafafa" stroke-width="1" opacity="0.7"/> <path d="M32 22 L22 32 L32 42 L42 32 Z" fill="none" stroke="#fafafa" stroke-width="1.2"/> <circle cx="32" cy="32" r="6" fill="none" stroke="#fafafa" stroke-width="0.8"/> <circle cx="32" cy="32" r="2.5" fill="#fafafa"/> </svg>

Akeel Thought Injection Dataset (50K Augmented)

Multi-turn augmented reasoning traces for robust thought injection training

A QRK Labs Research Dataset

</div>

Overview

This dataset contains 50,000 augmented examples designed to prevent overfitting when training thought injection models. It expands the base 20K dataset through:

1. 1.Shuffled originals (17,500) — Base samples randomly reordered
2. 2.System prompt variations (17,500) — Same Q&A with different instruction phrasings
3. 3.Multi-turn conversations (15,000) — 2-4 chained Q&A pairs per conversation

Why Augmentation?

Single-turn training can cause models to overfit to specific response patterns. This augmented dataset helps models:

• —Maintain context across multiple turns
• —Generalize across different system prompts
• —Learn robust thought injection behavior

Format

Single-turn (same as base dataset)

<|im_start|>system
You are a knowledgeable assistant...<|im_end|>
<|im_start|>user
Question here<|im_end|>
<|im_start|>assistant
<think>...</think>
Answer here<|im_end|>

Multi-turn (2-4 exchanges)

<|im_start|>system
You are an AI assistant with access to external knowledge...<|im_end|>
<|im_start|>user
First question<|im_end|>
<|im_start|>assistant
<think>...</think>
First answer<|im_end|>
<|im_start|>user
Second question<|im_end|>
<|im_start|>assistant
<think>...</think>
Second answer<|im_end|>

Statistics

Augmentation Details

System Prompt Variations

Five equivalent phrasings are used:

1. 1."You are a knowledgeable assistant that provides accurate, helpful answers..."
2. 2."You are an AI assistant with access to external knowledge..."
3. 3."You are a helpful assistant. When answering questions, use <knowledge> tags..."
4. 4."You are an intelligent assistant that retrieves information..."
5. 5."You are a knowledgeable AI. Request external information with <knowledge> tags..."

Multi-turn Construction

• —Randomly samples 2-4 independent Q&A pairs
• —Chains them into a single conversation
• —Assigns a random system prompt variation
• —Final shuffle ensures no ordering bias

Usage

from datasets import load_dataset

dataset = load_dataset("qrk-labs/akeel-thought-injection-50k-augmented")

Related

• —[akeel-thought-injection-20k](https://huggingface.co/datasets/qrk-labs/akeel-thought-injection-20k) — Base single-turn dataset
• —[akeel-cot](https://huggingface.co/qrk-labs/akeel-cot) — Research model trained on this data

License

Apache 2.0

<div align="center"> _{Built with ☁️ by QRK Labs} </div>