Get a free Gemini API key (for Corpus Lab & the final project)

Reproducible LLM results, still 100% inside Colab — no install

When you need this

You need a key starting on Day 3. Days 1 and 2 are keyless:

Day 1 uses Colab’s built-in Gemini (colab.ai) for your first live call — zero setup. That free backend gives you no control over randomness (run the same prompt twice and the answer can change); that’s fine here, and we use it to show you why LLM output varies.
Day 2 doesn’t call the model at all — it evaluates a frozen set of predictions committed to the repo, so everyone’s precision/recall/F1 come out identical while you’re learning the metrics.

From Day 3 on — the Corpus Lab and the final project — you run the model yourself, and your results have to be stable and reproducible: your reported precision/recall/F1 shouldn’t change every time you re-run, and (because the Corpus Labs are autograded) a grader must be able to re-run your notebook and get the same numbers. For that you switch to the Gemini API, where you can pin the randomness (temperature=0, a fixed seed, a pinned model).

Important

You do not install anything or leave Colab. You get a free key, paste it into Colab’s built-in Secrets manager once, and the notebook picks it up automatically.

Step 1 — Create a free API key (AI Studio)

Go to https://aistudio.google.com/apikey and sign in with your Tohoku Google account.
Click Create API key.
Copy the key (a long string starting with AIza…). Treat it like a password.

This is the free tier — no billing, no credit card. It has rate limits (a cap on requests per minute / per day), which are plenty for one student’s coursework.

Never paste your key into a notebook cell

Anyone who sees your notebook (or the committed .ipynb) would get your key. Always use Colab Secrets (Step 2) so the key stays out of the code and out of what you submit.

Step 2 — Add the key to Colab Secrets

Open your Corpus-Lab / project notebook in Colab.
In the left sidebar, click the 🔑 key icon (“Secrets”).
Click + Add new secret.
Name it exactly GEMINI_API_KEY.
Paste your key into the Value box.
Turn on the Notebook access toggle for this secret.

You only do this once per Google account — the secret is remembered across your notebooks.

Step 3 — The notebook uses it automatically

When GEMINI_API_KEY is set, the notebook’s setup cell reads it from Secrets and routes calls through the Gemini API instead of colab.ai. You don’t change any code. Under the hood it does this:

import os
from google.colab import userdata
from google import genai
from google.genai import types

# Pull the key from Colab Secrets (never hard-code it).
os.environ["GEMINI_API_KEY"] = userdata.get("GEMINI_API_KEY")

client = genai.Client()   # reads GEMINI_API_KEY from the environment

def generate_text(p):
    return client.models.generate_content(
        model="gemini-2.5-flash",                    # pinned model
        contents=p,
        config=types.GenerateContentConfig(
            temperature=0,   # greedy decoding — removes sampling randomness
            seed=42,         # fixed random state
        ),
    ).text

The setup cell prints which backend it chose, e.g. LLM backend: Gemini API (gemini-2.5-flash). If it still says Colab Gemini, your secret isn’t set or its notebook-access toggle is off — recheck Step 2.

About reproducibility (important, and a little humbling)

temperature=0 + seed makes the model as reproducible as a hosted LLM gets — but not perfectly. Requests are batched with other users’ on Google’s servers, and the resulting floating-point math can differ slightly run to run, which occasionally flips the label on a genuinely borderline item. So expect:

Clear-cut items → identical every run.
Ambiguous items → the same most of the time, with the odd flip.

Because of this, your final deliverable freezes your predictions: you run the model once, save its raw output to a JSON file, and commit that alongside your analysis. Your evaluation (P/R/F1, confusion matrix, error table) then runs off the frozen file, so your submitted numbers are 100% reproducible — and anyone can audit exactly what the model said. (The Corpus-Lab notebook shows you how.)

Rate limits & good manners

The free tier limits requests per minute and per day. If you hit a limit you’ll get a rate-limit error — wait a bit and re-run, or reduce your sample size. Practical tips:

Develop your prompt on a small slice (10–20 items) first; only run the full set once you’re happy.
Once you’ve frozen your predictions, you don’t need to call the API again — re-run analysis off the JSON.

Want to keep working after the course? (optional, advanced)

Everything above stays in Colab. If you’d rather run locally in your own environment (e.g. Positron + uv), clone the course repo and set the same GEMINI_API_KEY as an environment variable — the notebooks detect they’re not in Colab and use the API automatically. See Python setup (local). This is optional — not required for the course.

--- title: "Get a free Gemini API key (for Corpus Lab & the final project)" subtitle: "Reproducible LLM results, still 100% inside Colab — no install" toc: true --- ## When you need this **You need a key starting on Day 3.** Days 1 and 2 are keyless: - **Day 1** uses Colab's built-in Gemini (`colab.ai`) for your first live call — zero setup. That free backend gives you **no control over randomness** (run the same prompt twice and the answer can change); that's fine here, and we use it to *show* you why LLM output varies. - **Day 2** doesn't call the model at all — it evaluates a **frozen** set of predictions committed to the repo, so everyone's precision/recall/F1 come out identical while you're learning the metrics. **From Day 3 on — the Corpus Lab and the final project — you run the model yourself, and your results have to be stable and reproducible:** your reported precision/recall/F1 shouldn't change every time you re-run, and (because the Corpus Labs are **autograded**) a grader must be able to re-run your notebook and get the same numbers. For that you switch to the **Gemini API**, where you *can* pin the randomness (`temperature=0`, a fixed `seed`, a pinned model). ::: {.callout-important} You do **not** install anything or leave Colab. You get a free key, paste it into Colab's built-in **Secrets** manager once, and the notebook picks it up automatically. ::: ## Step 1 — Create a free API key (AI Studio) 1. Go to **<https://aistudio.google.com/apikey>** and sign in with your **Tohoku Google account**. 2. Click **Create API key**. 3. Copy the key (a long string starting with `AIza…`). Treat it like a password. This is the **free tier** — no billing, no credit card. It has rate limits (a cap on requests per minute / per day), which are plenty for one student's coursework. ::: {.callout-warning} ## Never paste your key into a notebook cell Anyone who sees your notebook (or the committed `.ipynb`) would get your key. Always use Colab **Secrets** (Step 2) so the key stays out of the code and out of what you submit. ::: ## Step 2 — Add the key to Colab Secrets 1. Open your Corpus-Lab / project notebook in Colab. 2. In the **left sidebar**, click the **🔑 key icon** (“Secrets”). 3. Click **+ Add new secret**. 4. Name it exactly `GEMINI_API_KEY`. 5. Paste your key into the **Value** box. 6. Turn on the **Notebook access** toggle for this secret. You only do this once per Google account — the secret is remembered across your notebooks. ## Step 3 — The notebook uses it automatically When `GEMINI_API_KEY` is set, the notebook's setup cell reads it from Secrets and routes calls through the Gemini API instead of `colab.ai`. You don't change any code. Under the hood it does this: ```python import os from google.colab import userdata from google import genai from google.genai import types # Pull the key from Colab Secrets (never hard-code it). os.environ["GEMINI_API_KEY"] = userdata.get("GEMINI_API_KEY") client = genai.Client() # reads GEMINI_API_KEY from the environment def generate_text(p): return client.models.generate_content( model="gemini-2.5-flash", # pinned model contents=p, config=types.GenerateContentConfig( temperature=0, # greedy decoding — removes sampling randomness seed=42, # fixed random state ), ).text ``` The setup cell prints which backend it chose, e.g. `LLM backend: Gemini API (gemini-2.5-flash)`. If it still says `Colab Gemini`, your secret isn't set or its notebook-access toggle is off — recheck Step 2. ## About reproducibility (important, and a little humbling) `temperature=0` + `seed` makes the model **as reproducible as a hosted LLM gets** — but **not perfectly**. Requests are batched with other users' on Google's servers, and the resulting floating-point math can differ slightly run to run, which occasionally flips the label on a genuinely borderline item. So expect: - **Clear-cut items** → identical every run. - **Ambiguous items** → the same *most of the time*, with the odd flip. Because of this, your final deliverable **freezes your predictions**: you run the model once, save its raw output to a JSON file, and commit that alongside your analysis. Your evaluation (P/R/F1, confusion matrix, error table) then runs off the **frozen file**, so your submitted numbers are 100% reproducible — and anyone can audit exactly what the model said. (The Corpus-Lab notebook shows you how.) ## Rate limits & good manners The free tier limits **requests per minute** and **per day**. If you hit a limit you'll get a rate-limit error — wait a bit and re-run, or reduce your sample size. Practical tips: - Develop your prompt on a **small** slice (10–20 items) first; only run the full set once you're happy. - Once you've frozen your predictions, you don't need to call the API again — re-run analysis off the JSON. ## Want to keep working after the course? (optional, advanced) Everything above stays in Colab. If you'd rather run locally in your own environment (e.g. Positron + `uv`), clone the course repo and set the same `GEMINI_API_KEY` as an environment variable — the notebooks detect they're *not* in Colab and use the API automatically. See [Python setup (local)](./python-setup.md). This is **optional** — not required for the course.