Datasets:
The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
chat-dataset-16b
A 16-billion-token, pre-tokenized, memory-mapped corpus engineered to pretrain ~400M–500M parameter chat models from scratch.
Unlike raw text corpora that produce standard auto-regressive paragraph completers, every document in this dataset is pre-formatted with ChatML tags and paired with an exact loss-masking label array (-100 on prompts, target token IDs on assistant responses). This teaches the model conversational turn structure and instruction following directly during pretraining.
Dataset Summary
- Total Tokens: 16,000,000,000 (16 Billion)
- Sequence Length: 2,048 tokens (fixed-window chunking)
- Total Sequences: ~7.81 million windows
- Split: 95% Train (
15.2B tokens) / 5% Validation (0.8B tokens) - Tokenizer:
HuggingFaceTB/SmolLM2-360M-Instruct(Vocabulary: 49,152) - Storage Format: Raw binary streams (
.bin) designed for zero-RAMnumpy.memmapreading
Corpus Composition
The dataset balances dense synthetic instruction, educational content, and code logic:
| Component | Share | Tokens | Source | Role |
|---|---|---|---|---|
| Education / Textbooks | 60% | ~9.6B | HuggingFaceTB/smollm-corpus (cosmopedia-v2) |
High-density textbook chapters, academic explanations, and summaries |
| Code & Reasoning | 25% | ~4.0B | HuggingFaceTB/smollm-corpus (python-edu) |
Python scripts, technical documentation, and structured algorithmic logic |
| Conversations | 15% | ~2.4B | HuggingFaceTB/smol-smoltalk |
Concise multi-turn dialogues curated for sub-1B architectures |
Data Structure & Format
Each split provides two parallel raw binary files:
- Downloads last month
- 28