metadata
license: cc0-1.0
language:
- en
task_categories:
- text-generation
pretty_name: CompactLM Synthetic Tasks 600
tags:
- synthetic
- instruction-following
- small-language-models
- demo
CompactLM Synthetic Tasks 600
A fully synthetic, unofficial demonstration instruction dataset for experimenting with text-processing pipelines and Hugging Face Datasets. Contains 600 template-based records spanning 20 AI engineering topics and 6 task categories (definition, glossary, summarization, rewriting, extraction, evaluation). There are 540 training records and 60 validation records.
Important limitations
- All entries were generated using deterministic templates, not sampled from people or real users.
- Responses are repetitive and have not been independently reviewed. This dataset is not a benchmark, a validated training corpus, or evidence of model performance.
- Train/validation splits can share closely related templates and subjects; do not treat validation scores as unbiased generalization evidence.
- Not affiliated with Hugging Face, NVIDIA, or another organization.
Format
Each JSONL object contains id, category, topic, messages (user and assistant turns), synthetic, source, and quality_note.
Example
{"id":"compactlm-0001","category":"definition","topic":"GPU inference","messages":[{"role":"user","content":"Explain GPU inference in one sentence for a beginner."},{"role":"assistant","content":"GPU inference means running a trained model on a graphics processor to produce predictions or outputs."}],"synthetic":true,"source":"template-generated","quality_note":"unreviewed demonstration sample"}
Loading
from datasets import load_dataset
dataset = load_dataset("json", data_files={"train": "train.jsonl", "validation": "validation.jsonl"})
print(dataset["train"][0])
Intended use
Education, data loading demos, schema testing, and experiments with instruction formatting. Not suitable as a sole training dataset for a production assistant.