Alpaca Cleaned Instruction Fine-Tuning Dataset

## Overview Alpaca-Cleaned is a quality-improved version of the Stanford Alpaca instruction-tuning dataset. It contains approximately 52,000 instruction-input-output triples in English, distributed as JSON, intended for supervised fine-tuning of instruction-following language models. The dataset is a static snapshot last updated April 2023. ## Schema - instruction — string — the task or question posed to the model - input — string — optional context or input data for the instruction (often empty) - output — string — the target response originally generated by GPT-3 (text-davinci-003), corrected during cleaning ## Sources - Original Stanford Alpaca: https://github.com/tatsu-lab/stanford_alpaca (CC-BY-NC 4.0 for the original; cleaned version released under CC-BY-4.0 by maintainer) - Cleaned version repo: https://github.com/gururise/AlpacaDataCleaned - HuggingFace mirror: https://huggingface.co/datasets/yahma/alpaca-cleaned (license: cc-by-4.0) ## Methodology The original Alpaca dataset was generated via self-instruct: seed tasks were expanded by GPT-3 (text-davinci-003) into ~52K instruction-following examples. The cleaned version applies systematic fixes to issues observed in the raw release, including: removing/repairing hallucinated answers (instructions referencing web data the model couldn't access), correcting merged instructions, filling in empty or N/A outputs, fixing wrong answers, removing instructions that required unsupported modalities (images, audio), and cleaning extraneous escape characters and formatting artifacts. ## Known gaps & limitations - English-only; not suitable for multilingual instruction tuning without augmentation. - Outputs originate from GPT-3 (text-davinci-003) and inherit that model's biases, factual errors, and stylistic patterns despite cleaning. - Cleaning was largely heuristic/manual; residual quality issues likely remain. - Not deduplicated against modern instruction benchmarks (MMLU, BBH, etc.) — possible contamination if used to train models later evaluated on those suites. - License of the original Stanford release is CC-BY-NC; downstream users relying on the CC-BY-4.0 relicensing of derivative work should validate their own legal posture for commercial use. ## Intended use & out-of-scope - IS for: supervised fine-tuning of instruction-following LLMs, LoRA/QLoRA experiments, baseline comparisons, ablation studies, and academic research on instruction tuning. - NOT for: production-grade factual QA without further filtering; multilingual training; benchmark-clean fine-tuning without contamination checks against common eval suites. Original supplier listing: Alpaca Cleaned — Instruction Fine-Tuning Dataset Cleaned version of Stanford's Alpaca instruction-following dataset (~52K examples). Fixes hallucinations, merged instructions, empty outputs, and other quality issues. CC-BY-4.0, ready for LLM fine-tuning.