Dhanishtha-2.0-SUPERTHINKER

📦 **Dhanishtha-2.0-SUPERTHINKER**

A distilled corpus of 11.7K high-quality samples showcasing multi-phase reasoning and structured emotional cognition. Sourced directly from the internal training data of **Dhanishtha-2.0** — the world’s first Large Language Model (LLM) to implement **Intermediate Thinking**, featuring multiple `<think>` and `<ser>` blocks per response

---

### 📊 Overview

* **11.7K multilingual samples** (languages listed below)
* **Instruction-Output** format, ideal for supervised fine-tuning
* **Recursive reasoning blocks** via `<think>...</think>`
* **Emotional state modeling** via `<ser>...</ser>`
* **Prompts include, but are not limited to,** philosophical, technical, personal, ethical, mathematical, and coding problems
* Data is derived from HelpingAI’s flagship LLM [Dhanishtha-2.0](https://huggingface.co/HelpingAI/Dhanishtha-2.0)

#### 🌐 Languages Included

Afrikaans, Arabic, Bulgarian, Catalan, Chinese (Simplified), Czech, Danish, Dutch, English, Estonian, Finnish, French, German, Greek, Hebrew, Hindi, Hungarian, Indonesian, Italian, Japanese, Korean, Marathi, Norwegian, Persian, Polish, Portuguese, Romanian, Russian, Spanish, Swahili, Swedish, Tagalog, Tamil, Telugu, Thai, Turkish, Ukrainian, Urdu, Vietnamese

---

### **⚙️ Intermediate Thinking: The Core Innovation**

**Unlike most LLMs that generate single-pass responses, Dhanishtha-2.0 engages in Intermediate Thinking—the ability to pause, reflect, and refine reasoning throughout its response.**

```plaintext
<think>
This is a physics question about gravitational forces. The Moon exerts a gravitational pull on objects on Earth. The direction of this force would be along the line connecting the object at the equator to the center of the Moon.
</think>

The direction of the gravitational force is along the line connecting the object to the Moon.

<think>
Wait, I should clarify that gravitational force follows Newton's law of universal gravitation. The direction is not static—it varies depending on the Moon's position relative to Earth.
</think>

So, to be more precise, the gravitational force from the Moon on an object at the equator points directly toward the Moon’s position in the sky at that time. As the Earth rotates and the Moon orbits, the direction of this force constantly shifts.

<think>
Maybe I should connect this to real-world effects like ocean tides.
</think>

This changing gravitational pull is what drives ocean tides. When the Moon is overhead, the pull is upward; when it's on the horizon, the pull is nearly horizontal.
```

**This enables:**

* **✅ Self-correction mid-response**
* **🔄 Question clarification and reframing**
* **📚 Multi-perspective synthesis**
* **🤖 Chain-of-self-correction, not just chain-of-thought**

---

### 💬 Structured Emotional Reasoning

The `<ser>...</ser>` block adds structured emotional awareness, encoding the model's contextual frame of mind:

```plaintext
<ser>
Emotion ==> curiosity
Cause ==> ambiguous ethical dilemma
Mind ==> conflicted but analytical
Growth ==> learning to navigate uncertainty
</ser>
```

**what does this mean?**

Each field in the `<ser>` block encodes a distinct aspect of the model's emotional and cognitive state during its response:

* **Emotion**: The model's inference about the user's current emotional state or the emotional tone of the prompt. This could include feelings such as curiosity, empathy, confusion, frustration, excitement, or skepticism. For example, if a user asks a philosophical question, the model might detect "curiosity" or "wonder."
* **Cause**: The underlying reason or trigger for the detected emotion. This is typically derived from the context of the prompt or the nature of the question. For instance, an ambiguous ethical dilemma might cause confusion or introspection, while a technical challenge could prompt determination or focus.
* **Mind**: The model's own cognitive stance or approach as it processes the prompt. This describes how the model is thinking—whether it is analytical, conflicted, cautious, optimistic, skeptical, or open-minded. For example, "conflicted but analytical" indicates the model is weighing multiple perspectives while maintaining logical rigor.
* **Growth**: The developmental or learning trajectory the model is experiencing as it reasons through the prompt. This field reflects how the model adapts, learns, or evolves its understanding in response to the user's input. Examples include "learning to navigate uncertainty," "becoming more empathetic," or "refining problem-solving strategies."

This structured emotional annotation enables the model to:

* Recognize and respond to the user's emotional context more effectively
* Provide transparency about its own reasoning and emotional stance
* Adapt its answers to be more empathetic, nuanced, and context-aware
* Handle complex, ambiguous, or sensitive prompts with greater care and self-reflection

---

### 🧠 Reasoning Workflow

Each example is composed of:

1. Prompt interpretation
2. Initial `<think>` (exploration)
3. Intermediate `<think>` (correction, roughwork)
4. Optional `<ser>` (emotional context)
5. Final answer (often revised or synthesized)

---

### 🧪 Data Collection & Processing

* **Origin**: All samples were extracted from the internal dataset used in training [Dhanishtha-2.0](https://huggingface.co/Abhaykoul/Dhanishtha-2.0), specifically focusing on the model’s Intermediate Thinking capabilities.
* **Filtering**: This dataset comprises a 0.25% uniformly random sample of the full Intermediate Thinking dataset used in Dhanishtha-2.0's training. No manual quality filtering was applied post-sampling, aside from ensuring structural integrity of reasoning blocks.
* **Languages**: The dataset includes examples in 39 languages, such as English, Hindi, Tamil, French, Arabic, Chinese (Simplified), Russian, and more. Samples may be monolingual or code-switched.
* **Validation**:

* Manual review of 1,000 randomly sampled entries to ensure syntactic correctness of `<think>` and `<ser>` tags
* Heuristic parsing to flag and discard malformed reasoning/emotion blocks
* Deduplication pass to eliminate near-identical prompt-response pairs

---

### 🧪 Sample Format

```json
{
"instruction": "What is the direction of the gravitational force exerted by the Moon on an object at the Earth's equator?",
"output": "<think>\nThis is a physics question about gravitational forces. The Moon exerts a gravitational pull on objects on Earth. The direction of this force would be along the line connecting the object at the equator to the center of the Moon.\n</think>\n\nThe direction of the gravitational force is along the line connecting the object to the Moon.\n\n<think>\nWait, I should clarify that gravitational force follows Newton's law of universal gravitation. The direction is not static—it varies depending on the Moon's position relative to Earth.\n</think>\n\nSo, to be more precise, the gravitational force from the Moon on an object at the equator points directly toward the Moon’s position in the sky at that time. As the Earth rotates and the Moon orbits, the direction of this force constantly shifts.\n\n<think>\nMaybe I should connect this to real-world effects like ocean tides.\n</think>\n\nThis changing gravitational pull is what drives ocean tides. When the Moon is overhead, the pull is upward; when it's on the horizon, the pull is nearly horizontal."
}
```

---

### 🔧 Quickstart

```python
from datasets import load_dataset

dataset = load_dataset("HelpingAI/Dhanishtha-2.0-SUPERTHINKER", split="train", streaming=True)

for row in dataset:
print(row["instruction"])
print(row["output"])
```

---

### 🔬 Intended Use

* Finetuning multilingual or emotionally aware LLMs
* Evaluating alignment under ethical and moral uncertainty
* Benchmarking multi-phase reasoning or internal monologue generation
* Training COT models that can self-correct or reflect on ambiguity

### 📄 Citation

```bibtex
@misc{HAI2025dhanishtha,
title = {Dhanishtha-2.0: A Large Language Model with Intermediate Thinking and Structured Empathy},
author = {Abhay Koul and Varun Gupta},
year = {2025},
publisher = {HelpingAI},
howpublished = {\url{https://huggingface.co/HelpingAI/Dhanishtha-2.0-SUPERTHINKER}}
}
```

---

🗭 *Dhanishtha-2.0 SUPERTHINKER isn't just data — it's a journey into recursive reflection and structured emotional cognition, empowering language models to reason, self-correct, and grow through multiple phases of Intermediate Thinking.*

📦 **Dhanishtha-2.0-SUPERTHINKER**

本数据集为经过精简的语料库，包含11.7K条高质量样本，展示多阶段推理与结构化情感认知。其数据直接源自**Dhanishtha-2.0**的内部训练数据——这是全球首款实现**中间思考（Intermediate Thinking）**的大语言模型（Large Language Model, LLM），每轮回复中包含多个`<think>`与`<ser>`模块。

---

### 📊 概述

* **11.7K条多语言样本**（语言列表见下文）
* **指令-输出（Instruction-Output）格式**，适用于有监督微调
* **递归推理模块（Recursive reasoning blocks）**，通过`<think>...</think>`实现
* **情感状态建模（Emotional state modeling）**，通过`<ser>...</ser>`实现
* **提示内容涵盖但不限于**哲学、技术、个人、伦理、数学与编码类问题
* 数据源自HelpingAI旗下旗舰大语言模型 [Dhanishtha-2.0](https://huggingface.co/HelpingAI/Dhanishtha-2.0)

#### 🌐 包含语言

南非荷兰语、阿拉伯语、保加利亚语、加泰罗尼亚语、简体中文、捷克语、丹麦语、荷兰语、英语、爱沙尼亚语、芬兰语、法语、德语、希腊语、希伯来语、印地语、匈牙利语、印尼语、意大利语、日语、韩语、马拉地语、挪威语、波斯语、波兰语、葡萄牙语、罗马尼亚语、俄语、西班牙语、斯瓦西里语、瑞典语、他加禄语、泰米尔语、泰卢固语、泰语、土耳其语、乌克兰语、乌尔都语、越南语

---

### ⚙️ 中间思考：核心创新

与大多数仅生成单轮响应的大语言模型不同，Dhanishtha-2.0具备**中间思考（Intermediate Thinking）**能力——即在生成响应的过程中暂停、反思并优化推理路径。

plaintext
<think>
This is a physics question about gravitational forces. The Moon exerts a gravitational pull on objects on Earth. The direction of this force would be along the line connecting the object at the equator to the center of the Moon.
</think>

The direction of the gravitational force is along the line connecting the object to the Moon.

<think>
Wait, I should clarify that gravitational force follows Newton's law of universal gravitation. The direction is not static—it varies depending on the Moon's position relative to Earth.
</think>

So, to be more precise, the gravitational force from the Moon on an object at the equator points directly toward the Moon’s position in the sky at that time. As the Earth rotates and the Moon orbits, the direction of this force constantly shifts.

<think>
Maybe I should connect this to real-world effects like ocean tides.
</think>

This changing gravitational pull is what drives ocean tides. When the Moon is overhead, the pull is upward; when it's on the horizon, the pull is nearly horizontal.

该能力可实现：
* ✅ 响应过程中自我修正
* 🔄 问题澄清与重构
* 📚 多视角综合
* 🤖 自我修正链，而非仅思维链（Chain-of-Thought）

---

### 💬 结构化情感推理

`<ser>...</ser>`模块用于实现结构化情感感知，对模型生成响应时的上下文心智状态进行编码：

plaintext
<ser>
Emotion ==> curiosity
Cause ==> ambiguous ethical dilemma
Mind ==> conflicted but analytical
Growth ==> learning to navigate uncertainty
</ser>

各字段含义如下：
* **情感（Emotion）**：模型对用户当前情感状态或提示情感基调的推断，可涵盖好奇、共情、困惑、沮丧、兴奋或怀疑等情绪。例如，若用户提出哲学问题，模型可能检测到“好奇”或“惊叹”。
* **触发原因（Cause）**：引发检测到的情绪的潜在根源或触发因素，通常源自提示上下文或问题本质。例如，模糊的伦理困境可能引发困惑或内省，而技术挑战则可能激发决心或专注。
* **心智状态（Mind）**：模型处理提示时自身的认知立场或思路，描述模型的思考方式——如分析性、矛盾性、谨慎性、乐观性、怀疑性或开放性。例如，“矛盾但兼具分析性”表示模型在权衡多重视角的同时保持逻辑严谨。
* **成长轨迹（Growth）**：模型在推理过程中经历的发展或学习路径，反映模型如何根据用户输入调整、学习或深化理解。示例包括“学习应对不确定性”“提升共情能力”或“优化问题解决策略”。

该结构化情感标注可帮助模型：
* 更有效地识别并响应用户的情感上下文
* 提升自身推理与情感立场的透明度
* 生成更具共情性、更细腻且贴合上下文的回复
* 更谨慎地处理复杂、模糊或敏感的提示，并进行自我反思

---

### 🧠 推理工作流

每个样本由以下部分组成：
1. 提示解读
2. 初始`<think>`模块（探索阶段）
3. 中间`<think>`模块（修正、草稿阶段）
4. 可选`<ser>`模块（情感上下文）
5. 最终答案（通常经过修订或综合）

---

### 🧪 数据收集与处理

* **数据来源**：所有样本均提取自Dhanishtha-2.0训练所用的内部数据集，重点聚焦模型的中间思考能力。
* **筛选规则**：本数据集是Dhanishtha-2.0训练所用完整中间思考数据集的0.25%均匀随机采样样本。除确保推理模块的结构完整性外，未进行额外人工质量筛选。
* **语言覆盖**：数据集包含39种语言的样本，如英语、印地语、泰米尔语、法语、阿拉伯语、简体中文、俄语等。样本可为单语或语码混合形式。
* **验证流程**：
* 对1000条随机采样条目进行人工审核，确保`<think>`与`<ser>`标签的句法正确性；
* 采用启发式解析方法标记并丢弃格式错误的推理/情感模块；
* 执行去重操作，消除近乎完全一致的提示-响应对。

---

### 🧪 样本格式

json
{
"instruction": "What is the direction of the gravitational force exerted by the Moon on an object at the Earth's equator?",
"output": "<think>
This is a physics question about gravitational forces. The Moon exerts a gravitational pull on objects on Earth. The direction of this force would be along the line connecting the object at the equator to the center of the Moon.
</think>

The direction of the gravitational force is along the line connecting the object to the Moon.

<think>
Wait, I should clarify that gravitational force follows Newton's law of universal gravitation. The direction is not static—it varies depending on the Moon's position relative to Earth.
</think>

So, to be more precise, the gravitational force from the Moon on an object at the equator points directly toward the Moon’s position in the sky at that time. As the Earth rotates and the Moon orbits, the direction of this force constantly shifts.

<think>
Maybe I should connect this to real-world effects like ocean tides.
</think>

This changing gravitational pull is what drives ocean tides. When the Moon is overhead, the pull is upward; when it's on the horizon, the pull is nearly horizontal."
}

---

### 🔧 快速入门

python
from datasets import load_dataset

dataset = load_dataset("HelpingAI/Dhanishtha-2.0-SUPERTHINKER", split="train", streaming=True)

for row in dataset:
print(row["instruction"])
print(row["output"])

---

### 🔬 预期用途

* 对多语言或具备情感感知能力的大语言模型进行微调
* 评估伦理与道德不确定性场景下的模型对齐效果
* 对多阶段推理或内部独白生成能力进行基准测试
* 训练可进行自我修正或针对模糊性进行反思的思维链（Chain-of-Thought, COT）模型

---

### 📄 引用

bibtex
@misc{HAI2025dhanishtha,
title = {Dhanishtha-2.0: A Large Language Model with Intermediate Thinking and Structured Empathy},
author = {Abhay Koul and Varun Gupta},
year = {2025},
publisher = {HelpingAI},
howpublished = {url{https://huggingface.co/HelpingAI/Dhanishtha-2.0-SUPERTHINKER}}
}

---

🗭 *Dhanishtha-2.0 SUPERTHINKER绝非普通数据集——它是一场递归反思与结构化情感认知的探索之旅，赋能语言模型依托多阶段中间思考完成推理、自我修正与能力成长。*