README.md

---
license: mit
task_categories:
- text-generation
language:
- en
tags:
- sft
- conversational
- instruction-tuning
- multi-genre
- math-reasoning
- humour
- genz
- agent
size_categories:
- 10K-100K
configs:
- config_name: math_reasoning
  data_files: math_reasoning/*.parquet
- config_name: humour_chat
  data_files: humour_chat/*.parquet
- config_name: merged_genz_chat
  data_files: merged_genz_chat/*.parquet
- config_name: agent_chat
  data_files: agent_chat/*.parquet
---

# Filthy-data-SFT

This is a highly curated, cleaned, and structurally normalized version of the **`Arko007/Filthy-data`** dataset. Every file across all genres has been meticulously mapped into a standard SFT conversational sequence.

## Strict Data Quality Filtering
To protect models during fine-tuning from learning corrupt or blank behaviors, we applied a strict **Data Quality Pipeline**:
- **No Empty Turns**: Any prompt/response containing empty text strings (`""`) was thoroughly stripped out.
- **Coherent Conversations**: Removed conversational turns with null or invalid roles.
- **Complete Conversational Loops**: Dropped any thread that didn't have at least one valid user message and assistant answer.

## Subsets & Genre Overview

All records in this repository are saved as high-performance **Parquet** files organized into subdirectory paths corresponding directly to their genres.

| Genre Subset | Cleaned Records | Description |
| :--- | :--- | :--- |
| **`math_reasoning`** | 20504 | Curated mathematical problems, reasoning lines, and step-by-step logic |
| **`humour_chat`** | 5017 | Funny, witty, and contextual dialogue streams |
| **`merged_genz_chat`** | 1190 | Unified and restructured slang/colloquial GenZ and extreme filthy conversations |
| **`agent_chat`** | 22333 | System actions, structured rules, and agentic workflows |

---

## Data Schema

Every split matches this uniform, nested conversational schema:
- **`messages`** (list of dicts):
  - **`role`** (string): Either `"user"` or `"assistant"`.
  - **`content`** (string): Dialogue payload.

### Sample Representation
```json
{
  "messages": [
    {
      "role": "user",
      "content": "Yo, what is the vibe today?"
    },
    {
      "role": "assistant",
      "content": "No cap, we are just cooling out and vibing!"
    }
  ]
}
```

---

## Quick Start

```python
from datasets import load_dataset

# Load specific subsets seamlessly
agent_dataset = load_dataset("Arko007/Filthy-data-SFT", "agent_chat")
genz_dataset = load_dataset("Arko007/Filthy-data-SFT", "merged_genz_chat")

print(genz_dataset["train"][0])
```