Skip to main content
Back
40 XP

Lesson 7 · 10 min

Reading a model card

Pick the right model — and stop guessing — by reading the card like an engineer.

What to look for

When you land on a model card on Hugging Face, Anthropic, or OpenAI, check these in order:

  1. Parameter count — 7B, 70B, 405B. Roughly tracks capability and inference cost.
  2. Context window — max input + output combined. 8k? 128k? 1M?
  3. Training data cutoff — how recent is its world knowledge?
  4. Tokenizer — affects how much your text costs.
  5. License — is it usable commercially? Llama, Apache 2.0, MIT, proprietary?
  6. Benchmarks — MMLU, GSM8K, HumanEval, MT-Bench. With salt — benchmarks game easily.
  7. Safety / alignment — RLHF? DPO? Tool-use trained?