Bias
A model is a compression of its corpus, so it inherits the corpus’s skews. Nobody injects bias — the model just learns the statistics of the text it was shown, faithfully.
Last lesson you cleaned a dataset. This lesson is about what that dataset contained — because a model can only mirror the text it’s built from. If a viewpoint is rare in the data, it’s rare in the model; if it’s dominant, the model treats it as the default.
So “bias” here isn’t a bug someone added. It’s the data’s statistics, learned faithfully. That’s an important reframing: the fix is never “remove the bias” — it’s understanding whose text the corpus is made of, and deciding what to do about it.
And here’s the part that ties back to this whole stage: every collection and cleaning choice is a value choice. What counts as quality, what gets deduplicated, which languages are worth crawling — each one quietly shapes who the model serves best.
Training text isn’t a neutral sample of humanity — it’s mostly English and mostly Global-North. Select a language to compare its share of training text with its share of the world’s people.
Illustrative shares, rounded — real figures vary by corpus, but the shape is robust: English dominates the text far beyond its share of speakers. This is also why, from the Tokens lesson, Spanish and other languages cost more tokens and are modelled less well — the case for local, own-data models.
Remember the quality filter from the Data lesson? It rewards a particular style — formal, edited, encyclopaedic prose. The same idea said plainly scores lower. Tap a pair and watch the exact same meaning get two different scores.
“The municipal council convened to evaluate proposals for improving public transportation across the district.”
“the council got together to look at ideas for making the buses better around here.”
Same meaning. The formal version scores higher only because it matches the filter’s house style, not because it’s more true or more useful.
Same illustrative scorer as the Data lesson. The point isn’t the exact numbers — it’s that “quality” is defined against a norm, and every norm includes some voices and excludes others.
There’s a subtler kind of skew. Models learn from co-occurrence — which words show up near which. If a corpus overwhelmingly pairs one job with one gender, or one place with one adjective, the model absorbs that association as if it were fact. You can’t see it in any single sentence; it lives in the statistics. Two lessons from now, in Embeddings, you’ll see these associations as directions in vector space — and measure them.
There’s no unbiased corpus to reach for — only choices, each with trade-offs. These are the main levers, and reasonable people disagree about when and how far to use them.
Up-sample under-represented languages and sources so they carry more weight. Helps coverage — but too much repetition of scarce data hurts quality, so it’s a balancing act.
Commission and digitise text for low-resource languages and communities instead of only crawling what’s already online. Slow and costly, but it fixes the source, not just the symptom.
Datasheets and data statements record what a corpus contains and omits, so downstream users know its blind spots. Transparency, not a cure.
Bias evaluations probe a trained model for specific disparities. Useful signal — but a benchmark only tests what its authors thought to ask.
None of these produces a “neutral” model, and that’s the honest takeaway: there is no view from nowhere. The goal isn’t a bias-free corpus — it’s making the choices visible and deliberate instead of accidental.
You’ve followed the data all the way in — sourced, tokenized, cleaned, and skewed. Next we open the thing it trains: the neural network that turns all those token numbers into predictions.