README.md

---
license: apache-2.0
task_categories:
- text-generation
- question-answering
language:
- en
- code
tags:
- pretraining
- tiny-slm
- smollm
- unsloth
- reasoning
- fineweb-edu
- finemath
- cosmopedia
- tinystories
- wikipedia
- dclm
size_categories:
- 10M<n<100M
configs:
- config_name: default
  data_files:
  - split: train
    path: "train/*.parquet"
  - split: validation
    path: "validation/*.parquet"
---

# 🚀 Ultra High-Quality Tiny SLM Pre-Training Corpus (<100GB)

A state-of-the-art, balanced **7-domain** pre-training dataset engineered specifically for **Small Language Models (Tiny SLMs: 50M – 2B parameters)** such as SmolLM2, SmolLM3, MobileLLM, Llama 3.2 1B, and custom architectures.

100% compatible with **Unsloth Studio**, **Unsloth AI**, **Hugging Face `datasets`**, and **PyTorch DataLoaders**.

---

## 📊 Dataset Statistics

- **Total Documents**: 20,066,075
  - **Train**: 19,663,898
  - **Validation**: 402,177
- **Total Tokens**: ~23.87 Billion tokens
  - **Train Tokens**: ~23.40B
  - **Validation Tokens**: ~0.47B
- **Format**: Zstandard Compressed Parquet

---

## 🔬 7-Domain Pre-Training Mixture

| Domain | Source Repository | Description & Quality Filter | Token Share |
| :--- | :--- | :--- | :--- |
| **Synthetic Textbooks** | `HuggingFaceTB/smollm-corpus` | `cosmopedia-v2` textbooks and topic deep-dives | **28.0%** |
| **Educational Web** | `HuggingFaceFW/fineweb-edu` | Sample 10BT (score >= 2.8 filter) | **23.5%** |
| **Step-by-Step Math** | `HuggingFaceTB/finemath` | `finemath-4plus` & `infiwebmath-4plus` (score 4+) | **19.5%** |
| **Clean Polyglot Code** | `codeparrot/github-code-clean` | Python, Rust, C++, C, Go, JS, TS, Java, SQL, Shell | **12.5%** |
| **General Knowledge** | `wikimedia/wikipedia` + `mlfoundations/dclm-baseline-1.0-parquet` | English Wikipedia (dense factual) & DCLM web corpus | **7.8%** |
| **Reasoning & Tool Traces** | `HuggingFaceTB/smoltalk2` | OpenThoughts3, Nemotron R1, and agentic tool traces | **6.5%** |
| **Stories & Narrative** | `roneneldan/TinyStories` | Synthetic coherent narrative, dialogue, & vocabulary | **2.1%** |

---

## 🦥 Quickstart: Unsloth Studio & Hugging Face

### 1. Load with Hugging Face Datasets
```python
from datasets import load_dataset

# Load Train Split
train_dataset = load_dataset("JustACluelessKidAtSchool/tiny-slm-pretraining-corpus", split="train")

# Load Validation Split
val_dataset = load_dataset("JustACluelessKidAtSchool/tiny-slm-pretraining-corpus", split="validation")
```

### 2. Pre-Train with Unsloth in 1 Command
```python
from unsloth import FastLanguageModel
from datasets import load_dataset
from transformers import TrainingArguments
from trl import SFTTrainer

# Load Tiny Model (e.g. SmolLM2-135M or Llama-3.2-1B)
model, tokenizer = FastLanguageModel.from_pretrained(
    model_name="HuggingFaceTB/SmolLM2-135M",
    max_seq_length=2048,
    load_in_4bit=False,
)

# Load Dataset from Hub
dataset = load_dataset("JustACluelessKidAtSchool/tiny-slm-pretraining-corpus", split="train")

# Train with Unsloth
trainer = SFTTrainer(
    model=model,
    tokenizer=tokenizer,
    train_dataset=dataset,
    dataset_text_field="text",
    max_seq_length=2048,
    args=TrainingArguments(
        output_dir="./tiny_slm_output",
        per_device_train_batch_size=4,
        gradient_accumulation_steps=8,
        max_steps=5000,
        learning_rate=5e-4,
        fp16=True,
        optim="adamw_8bit",
    ),
)
trainer.train()
```

### 3. In Unsloth Studio GUI
1. Set Dataset to: `JustACluelessKidAtSchool/tiny-slm-pretraining-corpus`
2. Text Column: `text`