AIAny
Icon for item

databricks-dolly-15k

A 15,000+ English instruction–response corpus for fine-tuning and evaluating LLM instruction-following behavior. Contains human-authored prompts and answers across categories (closed/open QA, summarization, extraction, classification, brainstorming) and is released under CC BY-SA 3.0.

Introduction

Human-written instruction data that is permissively licensed and practical for commercial fine-tuning remains rare. This corpus supplies over 15k prompt–response pairs authored by Databricks contributors across several InstructGPT-style categories, offering a compact, human-grounded resource for instruction tuning and synthetic data generation.

What Sets It Apart
  • Human-authored prompts and responses: created by thousands of Databricks employees rather than scraped or model-generated examples, which helps preserve natural instruction styles and realistic answers.
  • Multi-category coverage: includes closed QA (with Wikipedia context), open QA, summarization, information extraction, classification, brainstorming and creative writing, so models can be fine-tuned for diverse instruction behaviors.
  • Permissive, commercial-friendly license: released under CC BY-SA 3.0, allowing modification and commercial use with share-alike attribution.
  • Compact and practical: ~15k records make it feasible to use directly for small-scale instruction tuning or as seed examples for synthetic expansion (e.g., Self-Instruct style augmentation).
Who It's For and Trade-offs

Great fit if you need a small, human-curated instruction corpus for initial instruction-tuning experiments, prompt engineering, or generating few-shot seeds for synthetic data pipelines. Look elsewhere if you require very large-scale, high-diversity datasets or domain-specific proprietary content: the dataset is modest in size and contains Wikipedia-derived references for some tasks, so factual coverage and domain breadth are limited.

Where It Fits

Use this dataset as a human-grounded seed for instruction tuning, for benchmarking basic instruction-following capability, or as input examples for automated data-augmentation workflows. For production-grade LLMs that demand broader or more up-to-date knowledge, combine it with larger web-scale corpora or domain-specific datasets.

Information

  • Websitehuggingface.co
  • OrganizationsDatabricks, Inc.
  • Published date2023/04/11

Categories

More Items

Hugging Face

A 16 GB, 507-file PhD‑level cybersecurity knowledge base for training and evaluating security-focused LLMs and automation. Covers offensive/defensive/forensics/cloud/iot and AI-security across 30+ domains with real-world labs and framework mappings.

Hugging Face

Structured dataset for training and evaluating LLM agentic behavior: function-calling conversations, JSON-mode structured outputs, and extraction samples for teaching models to generate tool calls and strict structured responses. Includes single-turn and multi-turn scenarios across several configs.

Hugging Face

A multi-task English NLU benchmark for evaluating models across nine tasks (acceptability, sentiment, paraphrase, similarity, and various NLI setups), with a diagnostic evaluation set and an online leaderboard to compare generalization and transfer learning.