marketeam/FineWeb-Marketing-Annotations

FineWeb-Marketing Annotations

495,116 FineWeb documents scored 0–5 for marketing content quality by Gemma-3-27B-it. This is the annotation data behind marketeam/Fineweb-Classifier-Marketing.

Dataset Description

Each row is one document sampled from FineWeb, scored on a 0–5 marketing-quality scale by prompting Gemma-3-27B-it three independent times and aggregating the results. This is the same relationship as HuggingFaceFW/fineweb-edu-llama3-annotations is to the FineWeb-Edu classifier, applied to the marketing domain instead of education.

The dataset ships as two splits, matching exactly what marketeam/Fineweb-Classifier-Marketing was trained and evaluated on:

from datasets import load_dataset

train_ds = load_dataset("marketeam/FineWeb-Marketing-Annotations", split="train")       # 445,128 rows
val_ds = load_dataset("marketeam/FineWeb-Marketing-Annotations", split="validation")    # 49,988 rows

Schema

Column	Type	Description
id	string	FineWeb document ID
text	string	Full document text
dump	string	Source Common Crawl dump (e.g. CC-MAIN-2021-04) — 48 distinct dumps are represented
url	string	Source URL
raw_scores	list<int8>	The 3 individual scores parsed from Gemma's 3 independent completions for this document (null where a completion failed to parse)
score	float32	Aggregated score: majority vote if ≥2 of the 3 samples agree, otherwise the mean of the 3
int_score	int8	round(clip(score, 0, 5)) — integer bucket, kept for FineWeb-Edu-style backwards compatibility
data_split	string	"train" or "validation" — redundant with which split (see above) a row belongs to; kept so the value travels with the row if it's ever combined with other data (e.g. concatenate_datasets)

Note on data_split/split assignment: it's derived by matching each row's text against marketeam/Fineweb-Classifier-Marketing's original split files, since those files only carry text+score (no id). A small number of documents (~0.01%) share identical text with another row elsewhere in this dataset; for those, the assignment is a best-effort tie-break rather than a guaranteed exact match.

Annotation Methodology

This process mirrors the FineWeb-Edu annotation methodology, adapted to the marketing domain with refinements validated by JQL (Ali et al., 2025), a controlled study of LLM-annotator pipelines for this style of quality classifier.

Sampling: 500,000 documents were streamed from FineWeb's sample-350BT subset with a reproducible shuffle (buffer size 100,000, seed 42).

Why Gemma-3-27B-it: JQL evaluated 12 candidate annotator LLMs against human ground truth and found Gemma-3-27B-it achieved the highest Spearman correlation among them (JQL Fig. 2) — it was selected as the annotator on that basis, in place of the Llama-3-70B-Instruct used by the original FineWeb-Edu pipeline.

Prompting: a G-Eval-style prompt (per-criterion reasoning followed by a final score, with few-shot anchor examples) asks the model to work through the 5 rubric criteria below in order, one sentence of reasoning each, then sum to a 0–5 score. google/gemma-3-27b-it is prompted with temperature=1.0, top_k=64, top_p=0.95, min_p=0.0, generating 3 independent completions per document (document text truncated to 12,000 characters before prompting). Each completion is parsed for a line of the form Marketing score: <0-5> (last occurrence wins if the model repeats the label). Annotating the full 500,000-document sample was budgeted at roughly 400M input tokens (prompt + document text) and 360M output tokens (3 samples × ~240 reasoning tokens each).

What counts as marketing: the prompt defines a high-value marketing page as operating in at least one of five dimensions, with depth and rigor — marketing is treated as broader than copywriting:

Dimension	Description
T1 Knowledge	Marketing concepts, frameworks, theory, definitions (positioning, segmentation, JTBD, 4Ps, AIDA, CAC/LTV, attribution)
T2 Reasoning	Applying marketing logic to real situations: case studies, campaign teardowns, GTM analyses, positioning critiques, strategy memos
T3 Generation	Actual marketing creative: landing pages, ads, emails, product copy, brand voice, taglines, scripts, content marketing
T4 Analytics	Working with marketing data: A/B test writeups, attribution modeling, cohort and retention analyses, MMM, funnel diagnostics, dashboards, KPI methodology
T5 Operations	End-to-end marketing workflows: campaign planning, marketing automation, lifecycle programs, RevOps processes, ICP and outreach playbooks, lead routing, martech stack design

Rubric: each document is scored against 5 criteria, evaluated in order with one sentence of reasoning per criterion, then summed to a 0–5 score:

• C1 Relevance — genuinely about marketing in one or more of T1–T5 above (not just incidentally mentioning a product or brand)
• C2 Competence — specific rather than generic, demonstrates actual understanding of its subdomain, free of spam, fabricated claims, dark patterns, auto-generated SEO content, thin affiliate content, or boilerplate
• C3 Professional — coherent structure, consistent voice, would be recognized as legitimate work by a practitioner in that subdomain
• C4 Expert — correct application of frameworks, sound methodology, concrete data or examples, real strategic or analytical depth, non-obvious insights
• C5 Exceptional — best-in-class within its dimension: a definitive concept explainer, a teardown revealing new insight, copy that evidently drives action, an analytics writeup with rigorous methodology, or a workflow genuinely executable end-to-end

Guardrails built into the prompt to keep scoring from drifting toward superficial signals:

• Do not reward length — a short, sharp piece can score higher than a long padded one; SEO bloat is a C2 failure, not a C3 success.
• A page does not need to span multiple dimensions to score high — a pure T4 analytics writeup can score 5 if it's exceptional in T4.
• Technicality does not equal quality — a jargon-heavy page that says nothing concrete fails C2.
• Promotional intent alone is not a quality signal — a blunt sales page can pass C1 but fail C2.

Why 3 samples with majority-vote aggregation: a single stochastic LLM sample is a noisy label. JQL (§2.1, §3.1) found that sampling 3 independent completions per document and aggregating by majority vote — falling back to the mean when all 3 disagree — measurably reduces this label noise, at the cost of 3x annotation compute. This dataset keeps all 3 raw per-sample scores (raw_scores), not just the aggregate, so this aggregation step can be recomputed or studied directly.

Filtering: documents with fewer than 2 successfully parsed scores (out of 3) are dropped entirely. Of the 500,000 sampled documents, 495,116 (99.02%) retained a valid aggregated score — this dataset contains only those 495,116.

Used By

This is the full annotation set behind marketeam/Fineweb-Classifier-Marketing, which trains a regression head on top of a frozen encoder to predict this same 0–5 marketing-quality score.

Gemma-3-27B-it is the annotation model that produced these labels — it is not part of the classifier's own weights (the classifier's frozen encoder is Snowflake/snowflake-arctic-embed-m-v2.0; Gemma's role here is purely as a data-labeling tool during dataset construction).

Applications

• Training quality classifiers — the intended use, as demonstrated by marketeam/Fineweb-Classifier-Marketing.
• Pretraining corpus curation research — study percentile- or threshold-based filtering strategies for marketing-domain data, in the same spirit as FineWeb-Edu's use of its own annotations.
• LLM-as-judge research — the raw_scores column preserves all 3 individual samples (not just the aggregate), enabling study of inter-sample agreement, annotator calibration, and aggregation strategies.

Source & License

• Text source: HuggingFaceFW/fineweb, licensed ODC-By 1.0. This dataset is released under the same license, consistent with that requirement.
• Annotation model: Gemma-3-27B-it (google/gemma-3-27b-it).

Limitations

• Gemma-calibrated. Scores reflect Gemma-3-27B-it's judgment against the rubric in this card, not an objective or universal notion of marketing quality. A different annotator model would produce a different distribution.
• English-only. Sampled and annotated as English-language content only.
• Left-skewed distribution. As with any random web-crawl sample, most documents score low; high scores (4–5) are comparatively rare.
• Single-annotator-model labels. Each document received 3 samples from one model (Gemma-3-27B-it), not multiple independent annotator models or human review.