VINAY-UMRETHE/Sangam

---
dataset_info:
- config_name: full
features:
- name: messages
list:
- name: role
dtype: string
- name: content
dtype: string
- name: reasoning
dtype: string
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 603557645
num_examples: 70202
download_size: 412213960
dataset_size: 603557645
- config_name: non_reasoning
features:
- name: messages
list:
- name: role
dtype: string
- name: content
dtype: string
- name: reasoning
dtype: string
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 90963302
num_examples: 19099
download_size: 86823232
dataset_size: 90963302
- config_name: reasoning
features:
- name: messages
list:
- name: role
dtype: string
- name: content
dtype: string
- name: reasoning
dtype: string
- name: metadata
dtype: string
splits:
- name: train
num_bytes: 512594343
num_examples: 51103
download_size: 325438883
dataset_size: 512594343
configs:
- config_name: full
data_files:
- split: train
path: full/train-*
- config_name: non_reasoning
data_files:
- split: train
path: non_reasoning/train-*
- config_name: reasoning
data_files:
- split: train
path: reasoning/train-*
license: apache-2.0
task_categories:
- text-generation
language:
- en
tags:
- sangam
- distillation
- agent
- code
- math
- reasoning
- non_reasoning
- synthetic
---

# Sangam: From Pinnacles of Frontiers

<p align="center">
<img src="https://img.shields.io/badge/Size-70.2K-green?style=for-the-badge">
<img src="https://img.shields.io/badge/Dataset-Sangam-blue?style=for-the-badge">
<img src="https://img.shields.io/badge/License-Apache%202.0-orange?style=for-the-badge">
<img src="https://img.shields.io/badge/Format-OpenAI%20JSONL-red?style=for-the-badge">
</p>

**Sangam** is a multi-source instruction and reasoning dataset designed specifically for training and distilling large language models (LLMs) to exhibit advanced Chain-of-Thought (CoT), Agentic, Mathematical and Coding capabilities. It aggregates high-quality outputs from frontier models into a strict OpenAI `messages` format.

## Dataset Structure
The dataset contains a total of **70.2K** examples, split into three distinct subsets based on the presence of visible reasoning traces:

| Subset Name | Example Count | Description |
|-----------------|---------------|-------------|
| `full` | 70.2K | The globally merged dataset containing both reasoning and non-reasoning examples. |
| `reasoning` | 51.1K | Examples strictly containing a `<think>...</think>` block. The thinking has been isolated from the final assistant response. |
| `non_reasoning` | 19.1K | Purely conversational examples curated without explicit reasoning. |

---

### Schema

Every example rigidly conforms to the following schema structure:

```json
{
"messages": [
{"role": "user", "content": "string"},
{"role": "assistant", "content": "string"}
],
"reasoning": "string | null",
"metadata": "string" // mixed metadata
}
```

> **Note**: Because **Sangam** merges vastly diverse datasets (ranging from tightly-controlled synthetic outputs to raw API scraping containing deeply nested array of usage statistics, tokens, etc.), the underlying metadata structures are highly heterogeneous. As a result, the metadata is stored as a JSON string.

---

### All Possible Metadata Fields

> **Note**: Not all examples contain every field. Use JSON parsing to gracefully extract available keys.

```json
{
"source": "null | string",
"original_index": "integer",
"id": "string",
"difficulty": "string",
"category": "null | string",
"timestamp": "null | string",
"hash": "string",
"uuid": "string",
"domain": "string",
"meta": {
"cycle": "integer | null",
"original_difficulty": "null | string",
"sampling_temperature": "float",
"source_file": "string",
"teacher_model": "string",
"timestamp": "null | string",
"training_stage": "string"
},
"prompt": "null | string",
"response": "null",
"model": "null | string",
"chat_number": "null",
"response_in_chat": "null",
"concept": "string",
"text": "string",
"tools": {
"array_of": {
"type": "string",
"function": {
"name": "string",
"description": "string",
"parameters": {
"type": "string",
"properties": {
"file_path": {
"type": "string",
"description": "string"
},
"content": {
"type": "string",
"description": "string"
},
"old_text": {
"type": "string",
"description": "string"
},
"new_text": {
"type": "string",
"description": "string"
},
"dir_path": {
"type": "string",
"description": "string"
},
"pattern": {
"type": "string",
"description": "string"
},
"file_pattern": {
"type": "string",
"description": "string"
},
"command": {
"type": "string",
"description": "string"
},
"query": {
"type": "string",
"description": "string"
}
},
"required": "array[string]"
}
}
}
},
"metadata": {
"session_id": "string",
"turns": "integer",
"completed": "boolean",
"tool_calls_count": "integer",
"error": "null | string",
"source_dataset": "string",
"domain": "string",
"category": "null",
"model": "null | string",
"teacher_model": "string",
"original_difficulty": "null | string",
"uuid": "string",
"chat_number": "null",
"response_in_chat": "null",
"type": "string",
"difficulty": "string"
},
"usage": {
"prompt_tokens": "integer",
"completion_tokens": "integer",
"total_tokens": "integer",
"cost": "float"
},
"model_version": "string"
}
```
---

## Usage

### Loading the Dataset

```bash
pip install datasets
```
Sangam supports three configurations: `full`, `reasoning`, and `non_reasoning`.

```python
from datasets import load_dataset

# Load a subset
dataset = load_dataset("VINAY-UMRETHE/Sangam", "reasoning", split="train")
# dataset = load_dataset("VINAY-UMRETHE/Sangam", "non_reasoning", split="train")
# dataset = load_dataset("VINAY-UMRETHE/Sangam", "full", split="train")

print(dataset[0])
```

### Parsing Metadata

Since the `metadata` field is stored as a JSON string to ensure schema stability, you should parse it back into a dictionary if you need to access specific fields:

```python
import json

def parse_metadata(example):
example["metadata_dict"] = json.loads(example["metadata"])
return example

# Map the dataset to include a parsed metadata dictionary
dataset = dataset.map(parse_metadata)

print(dataset[0]["metadata_dict"]["source"])
```

### Training with Reasoning

```python
def format_for_reasoning(example):
# Extract the assistant message
messages = example["messages"]
# Append reasoning to the assistant's content if it exists
if example["reasoning"]:
for msg in messages:
if msg["role"] == "assistant":
msg["content"] = example["reasoning"] + "\n" + msg["content"]
return {"formatted_messages": messages}

dataset = dataset.map(format_for_reasoning)
```

---

## Models Utilized

Data within **Sangam** is sourced and unified from a variety of state-of-the-art model generations containing high-reasoning, agentic, Mathematical and coding capabilities:

| Model | Versions & Capabilities |
| :--- | :--- |
| **DeepSeek** | V3.2 (Speciale, Math) |
| **Claude** | 4.5 & 4.6 Opus & Sonnet (High-Reasoning, Writing-Style) |
| **Gemini** | 3.0 Pro & 3.1 Pro (High-Reasoning) |
| **GPT** | 5.0, 5.1 & 5.2 (Codex-Max, High) |
| **GLM** | 4.7 |
| **MiniMax** | M2.1 |

---
## License

**Sangam** is licensed under the **Apache 2.0** license. As mostly all sources were licensed under Apache or MIT.

---

## Citation

If you use this dataset in your research, please cite it as follows:

```bibtex
@misc{vinayumrethesangam2026,
author = {Vinay Umrethe},
title = {Sangam: From Pinnacles of Frontiers},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/VINAY-UMRETHE/Sangam}}
}
```