Rodrigo-Martin/productivity_enhancer_common_pile

Common Pile YouTube Productivity Categories

6,000 English-language-screened video titles sampled from Common Pile's YouTube collection and assigned title-only productivity-category labels using a fixed rubric.

These are AI-generated weak labels The release is a curated, balanced set of relatively clear examples; it does not represent natural YouTube class proportions.

Labels

Integer	Name	Operational definition
0	Productive	Science/technical education: the main purpose teaches or explains mathematics/statistics, computing/programming, engineering/electronics, natural sciences, or medical science.
1	Educational	The main purpose teaches a skill or explains another subject, including history, languages, cooking, music lessons, fitness, personal finance, psychology, law, politics, religion, or game strategy.
2	Unproductive	Primarily entertainment, performance, promotion, personal updates, gameplay, or reporting events without substantive instruction or explanation.

Splits

Split	Rows	Label 0	Label 1	Label 2
train	4,800	1,600	1,600	1,600
validation	600	200	200	200
test	600	200	200	200

Channels are assigned to one split by a seeded hash of channel ID. No channel is shared across splits. There are no repeated video IDs or case-folded, whitespace- normalized titles. At most 40 released examples come from any one channel.

Columns

Column	Type	Meaning
video_title	string	Source title, with whitespace normalized. This is the annotation input.
video_description	string	Original description retained for future experiments; not used to assign labels.
productivity_label	int64	Numeric class: 0, 1, or 2.
video_id	string	Original YouTube video ID.
channel_id	string	Original channel ID, also used to prevent split overlap.
source_url	string	Original video URL for provenance and attribution.
license	string	Source record's license entry.

Annotation and sampling

• Annotator: Qwen3-4B-Instruct-2507, MLX 4-bit conversion, run locally with greedy single-character output.
• Only the title is shown to the annotator. Creator identity, descriptions, transcripts, and existing labels are not provided.
• Ambiguous predictions, malformed outputs, generic titles without a subject, and possible class/purpose contradictions are excluded rather than forced into a class. A language detector screens longer titles for English.

Use

from datasets import load_dataset

dataset = load_dataset("Rodrigo-Martin/productivity_enhancer_common_pile")
train = dataset["train"]
validation = dataset["validation"]
test = dataset["test"]

Limitations

An AI annotator can misread a title or apply a boundary inconsistently. A title cannot establish a video's actual content, factual accuracy, quality, or usefulness to a particular person. Reviews and calibration examples do not establish a human-measured annotation accuracy for the release.

Common Pile's openly licensed, speech-based source selection and this release's clear-title filters introduce sampling biases. Short, vague, multilingual, or mixed- purpose videos are underrepresented. Balanced classes also differ from a typical recommendation feed.

Attribution

Upstream: Common Pile YouTube, part of The Common Pile v0.1, Kandpal et al., 2025. The source authors selected videos represented as Creative Commons Attribution content. This release retains the per-record license and source URL, together with channel IDs identifying the original creators. Original creators retain their rights in the underlying videos and descriptions.