README.md

---
license: apache-2.0
language:
- en
task_categories:
- text-generation
- conversational
- question-answering
tags:
- flatbuild
- conversational
- chatbot
- synthetic
- instruction
- dialogue
- training
- llm
- pretraining
size_categories:
- 10K<n<100K
---

# FlatBuild Demo Chat 10K Dataset

The **FlatBuild Demo Chat 10K Dataset** is the official conversational training dataset for **FlatBuild** and is used to train **Flatbot-Mini-35M**, the flagship demonstration language model of the **Flatseek** ecosystem.

The dataset showcases the complete workflow of building a conversational language model entirely from scratch, including:

- dataset preparation
- tokenizer training
- chat data preprocessing
- Transformer training
- checkpoint export
- GGUF conversion
- efficient inference with FlatRun

Designed for reproducibility and fast experimentation, the dataset enables developers to train compact language models on consumer hardware while demonstrating the complete FlatBuild pipeline.

---

# Overview

The dataset contains approximately **10,000** synthetic conversational examples covering a broad range of general-purpose assistant interactions.

Rather than focusing solely on instruction-following, the conversations emphasize natural dialogue, contextual understanding, and realistic multi-turn interactions.

The resulting models learn conversational capabilities such as:

- greetings
- introductions
- question answering
- explanations
- recommendations
- comparisons
- coding assistance
- troubleshooting
- brainstorming
- summarization
- simple reasoning
- follow-up conversations
- polite refusals
- general knowledge

---

# Dataset Format

Each line is stored as a JSON object representing a single conversation.

Example:

```json
{
  "messages": [
    {
      "role": "system",
      "content": "<system instruction>"
    },
    {
      "role": "user",
      "content": "<user message>"
    },
    {
      "role": "assistant",
      "content": "<assistant response>"
    }
  ]
}
```

Each conversation consists of an ordered sequence of chat messages using the standard `system`, `user`, and `assistant` roles. Conversation lengths range from single-turn exchanges to longer multi-turn dialogues.

---

# Characteristics

| Property | Value |
|----------|------:|
| Conversations | ~10,000 |
| Format | JSONL |
| Structure | Multi-turn chat |
| Language | English |
| Domain | General-purpose assistant |
| Data Type | Synthetic |
| Training Split | 95% |
| Validation Split | 5% |
| Recommended Context Length | 512 tokens |

---

# Topics

The dataset includes conversations across a diverse set of everyday subjects, including:

- greetings
- introductions
- daily life
- education
- productivity
- mathematics
- programming
- Python
- JavaScript
- technology
- artificial intelligence
- science
- history
- geography
- travel
- food
- weather
- books
- movies
- health
- finance
- recommendations
- comparisons
- troubleshooting
- brainstorming
- summarization
- logical reasoning
- general knowledge

Conversation lengths intentionally vary to encourage both short-response generation and longer context retention.

---

# Intended Use

This dataset is intended for:

- training conversational language models from scratch
- tokenizer training
- compact LLM research
- Transformer architecture experiments
- educational purposes
- reproducible language model training
- FlatBuild demonstrations

It is particularly suitable for developers interested in understanding how modern decoder-only language models can be trained without relying on pretrained foundation models.

---

# Not Intended For

This dataset is **not** intended for:

- production chatbots
- factual knowledge benchmarks
- safety evaluations
- alignment research
- instruction tuning of large pretrained models
- replacing large public conversational datasets

---

# Training

The dataset is used by the official FlatBuild training configuration for **Flatbot-Mini-35M**.

```bash
flatbuild train configs/flatbot-mini-35M.yaml
```

After training, checkpoints can be exported to SafeTensors or GGUF for inference with FlatRun or other GGUF-compatible runtimes.

---

# Design Goals

This dataset was created with the following objectives:

- Natural conversational flow
- Diverse wording
- Broad topic coverage
- Strong multi-turn context retention
- Minimal template repetition
- High-quality synthetic dialogue
- Fast training on consumer hardware
- Fully reproducible end-to-end workflow

Rather than maximizing dataset size, the focus is on producing a clean, high-quality conversational corpus that enables compact language models to learn practical dialogue behaviors efficiently.

---

# Models Trained Using This Dataset

This dataset is used by the official FlatBuild demonstration models, including:

- **Flatbot-Micro-4M**
- **Flatbot-Mini-35M**

Both models are trained entirely from random initialization using FlatBuild without relying on pretrained language models.

---

# License

Apache-2.0