Every species has a Pokédex-style entry. The base stats run from 0 to 100 and are ordinal and approximate: they are for comparing species, not measuring them. The first three (number of parameters, amount of data in examples or tokens, and training compute) use a log scale.
| Stat | What it measures | Anchors |
|---|---|---|
| Parameters | adjustable numbers | 15 ≈ 10³ · 45 ≈ 10⁸ · 75 ≈ 10¹¹ · 90 ≈ 10¹² |
| Data | examples or tokens | 10 = a spreadsheet · 45 ≈ 10⁸ · 90 ≈ 10¹³ |
| Compute | training FLOPs | 5 = pencil · 45 ≈ 10¹⁹ · 90 ≈ 10²⁵ |
| Context | how much input it considers at once | 0 = a single vector · 25 ≈ 512 tokens · 80 ≈ 128k |
| Generality | range of tasks | 10 = one · 40 = many after fine-tuning · 70 = many by prompting |
| Transparency | how readable it is | 95 = small tree · 60 = linear model · 10 = black box |
The data sheet adds hard numbers. Height is the number of layers: BERT-base stands 12 tall (Devlin et al., 2019, p. 4173). Weight is the memory the parameters take up: in fp32 precision each takes 4 bytes, so BERT-base's 110 million weigh ≈ 440 MB and GPT-3's 175 billion ≈ 700 GB (Brown et al., 2020, sec. 2.1); in fp16, half that, and less still with quantization.
Types (one or two per species, such as Neural or Recurrent) indicate the family of ideas it belongs to. The kind distinguishes architecture, model, algorithm, system, or unit (see model). Moves are what it can do; abilities, its distinctive architectural features; weaknesses, its known limits. Habitat says what data it grew up on and where it can be found today. The evolution chain links it to its ancestors and says what changed. Anatomy is a diagram of its parts, read from bottom to top, like the original transformer figure. The professor's note opens a philosophical discussion to take to class.
The stats are a useful caricature: a parameter count does not say how many are used at once (mixture of experts), and transparency is not a simple property (interpretability).