Datasets:

suehuynh
/

marketing-instruct-13k

license: apache-2.0
language:
  - en
tags:
  - marketing
  - instruction-tuning
  - copywriting
  - advertising
  - email-marketing
  - product-description
  - brand-voice
  - autoscientist
task_categories:
  - text-generation
size_categories:
  - 10K<n<100K

Dataset Card for marketing-instruct-13k

Dataset Details

Dataset Description

A curated instruction-tuning dataset of ~13,000 marketing copywriting examples across five task types, built for the AutoScientist Challenge 2026 (Marketing category). This is the largest dataset in the Marketing-Mixtral/Llama series, used to fine-tune Marketing-Llama-3.3-70B (88% win rate against base Llama-3.3-70B-Instruct).

It extends marketing-instruct-8k with an additional ~4,000 product description examples, and traces back to the original marketing-instruct-4k, the first dataset in this series.

Curated by: Sue Huynh
Language(s): English
License: Apache 2.0

Dataset Sources

Models trained on this data: Marketing-Llama-3.3-70B (88% win rate)
Related datasets: marketing-instruct-4k, marketing-instruct-8k
Challenge: AutoScientist Challenge 2026 — adaptionlabs.ai

Uses

Direct Use

Instruction fine-tuning for marketing copy generation models across: ad/social copy, email marketing, product descriptions, brand voice rewriting, and campaign data-to-insight summarization.

Out-of-Scope Use

Not designed for marketing strategy, competitive analysis, or market research generation tasks — these were explicitly excluded during curation to keep the dataset focused on copy generation
Not suitable for non-English marketing content without translation and re-validation
Not intended as a source of factual product information; product attributes used in examples are illustrative or sourced from public product listings, not independently verified

Dataset Structure

Each example follows a unified instruction-tuning schema:

{
  "instruction": "Write a product description for an online store listing...",
  "input": "",
  "output": "The Power Bank Model-39 packs 20000mAh..."
}

Field	Type	Description
instruction	string	The task brief, including format and constraints
input	string	Optional structured context (e.g. attribute list, table); empty string if not applicable
output	string	The target marketing copy

Task Type Distribution

Task Type	Approx. Examples	Notes
Product Description	~4,800	Attribute-grounded; largest task type in this version
Ad/Social Copy	~2,800	Includes Instagram-specific ad copy and brand voice pairs
Email Marketing	~2,400	Campaign, lifecycle, and personalized/persona-driven emails
Brand Voice Rewriting	~1,500	Neutral copy rewritten to match a specified voice
Table-to-Insight	~1,500	Campaign performance data summarized into insights

(Counts are approximate; see the dataset viewer for exact splits.)

Dataset Creation

Curation Rationale

This version extends the original curated dataset (marketing-instruct-4k) with additional public sources per task type, aiming to test whether broader source diversity and a larger product description set would improve fine-tuning results beyond the original curated set. It was trained alongside a smaller 8K-row sibling dataset (marketing-instruct-8k) to compare outcomes across dataset sizes and base models.

As with earlier versions, strategy/analysis tasks (SWOT analysis, competitor comparison tables, dashboards, market research) were explicitly excluded to keep the dataset focused on copy generation rather than business consulting.

Source Data

Data Collection and Processing

This dataset combines nine sources across three generations of curation:

Original sources (marketing-instruct-4k):

Ad copy generation — smangrul/ad-copy-generation (HuggingFace), adapted and quality-filtered via Adaption Labs' AdaptData platform
Brand voice briefs — manifesta/brandvoice-marketing-briefs (HuggingFace), prompt-only source; completions generated via AdaptData
Marketing copy prompts — marketeam/marketing_user_prompts_unfiltered (HuggingFace), filtered via SQL query to retain only copy-generation tasks, excluding strategy/analysis tasks; completions generated via AdaptData
Promotional email data — jason23322/high-accuracy-email-classifier (HuggingFace), filtered to the "promotions" category only (length ≥ 100 characters)
Synthetic examples — hand-designed across 40+ brand voice styles, 15+ industries, and varied table/insight scenarios, generated with AI assistance to fill gaps where public data was sparse or low quality

Expansion sources (marketing-instruct-8k): 6. Campaign performance data — manishabhatt22/marketing-campaign-performance-dataset (Kaggle), adapted via AdaptData to generate table-to-insight prompts and summaries 7. Personalized email data — LightTai/personalized-email (HuggingFace), adapted via AdaptData to generate prompts and customer personas 8. Marketing email samples — Isotonic/marketing_email_samples (HuggingFace), adapted via AdaptData to generate prompts and customer personas 9. Instagram social/ad copy — avihayamor/social-instagram-marketing (HuggingFace), adapted via AdaptData to generate ad copy prompts and brand voice pairs

This version's addition: 10. Amazon product descriptions — Ateeqq/Amazon-Product-Description (HuggingFace), adapted via AdaptData to add ~4,000 additional attribute-grounded product description examples

Processing steps applied across all sources:

Deduplication on output text hash
Removal of examples under 10 words (insufficient training signal)
Removal of placeholder-variable content (e.g. unfilled {time}, {items} tokens)
Removal of broken-encoding and non-UTF8 artifacts
Manual spot-checking of AdaptData-generated prompts/personas for quality and groundedness before merging

Who are the source data producers?

A mix of original dataset authors on HuggingFace and Kaggle (see individual source dataset cards for attribution), AdaptData's automated adaptation system (used to generate prompts, personas, and completions for several sources), and manual synthetic data authored for this project.

Personal and Sensitive Information

This dataset does not contain real personal, sensitive, or private information. Customer personas generated via AdaptData for the email marketing sources are synthetic constructs (e.g. "budget-conscious parent, age 30-40") and do not represent real individuals. Business names referenced in examples are either genericized, fictional, or based on publicly available product listings with manufacturer specifications only.

Bias, Risks, and Limitations

Skews toward Western, English-language marketing conventions and consumer/B2B SaaS contexts common in US-based marketing
Product description examples are now the largest task type in this version (~37% of the dataset), which may bias a model trained on this data toward stronger product description performance relative to other task types — this imbalance was not present in the smaller, more evenly distributed marketing-instruct-4k version
AdaptData-generated personas and prompts (sources 6-10) underwent spot-checking but not the same exhaustive faithfulness filtering applied to the original hand-curated product description examples
Synthetic examples may carry stylistic patterns from the language model used to assist in their creation

Recommendations

Models trained on this dataset should have their outputs reviewed by a human marketer before publication, particularly for factual product claims. Given the product description task type imbalance noted above, users seeking balanced performance across all five task types may prefer marketing-instruct-4k or marketing-instruct-8k.

Citation

BibTeX:

@misc{huynh2026marketinginstruct13k,
  author = {Huynh, Nguyen},
  title = {marketing-instruct-13k: An Extended Marketing Copywriting Instruction Dataset},
  year = {2026},
  publisher = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/suehuynh/marketing-instruct-13k}},
  note = {AutoScientist Challenge 2026 — Marketing Category}
}

Dataset Card Authors

Sue Huynh — Brown University MSc Data Science

Dataset Card Contact

HuggingFace: suehuynh