Choosing Evaluation Metrics for a New Model

8/15/2026beginner

Speakers

You

あなた

Yerlan

MLOpsエンジニア（カザフスタン）

この場面について

海外のMLOpsエンジニアと、新しいAIモデルの評価指標について打ち合わせる場面

Summary (JP)

あなたは、カザフスタン出身のMLOpsエンジニアYerlanとビデオ通話で、新しいAIモデルの評価方法について話し合います。二人は正確さだけでなく、応答速度や1リクエストあたりのコストも評価指標に加えることに合意します。テスト用データセットとして先月のユーザーの問い合わせデータを使うことになり、金曜日に結果を比較する約束をします。

会話

Speed:

文をクリックすると、その部分から読み上げが始まります。

You:

Hi Yerlan! Do you have a few minutes to talk about the new model?

Yerlan:

Sure! I'm free until three. What's on your mind?

You:

We need to choose evaluation metrics for it. Do you have any ideas?

Yerlan:

I think accuracy alone isn't enough this time. We should also look at speed.

You:

That makes sense. Users care about fast responses too.

Yerlan:

Right. I suggest we track accuracy, latency, and cost per request.

You:

Good idea. Where can we find a good test dataset?

Yerlan:

We can use last month's user queries. I already cleaned that data.

You:

Perfect. Can you share the dataset with me today?

Yerlan:

Of course. I'll send you the link after this call.

You:

Thanks! I'll run the first test tomorrow morning.

Yerlan:

Sounds good. Let's compare results on Friday.

You:

Agreed. I'll also write a short report for the team.

Yerlan:

Great. Let me know if you need help with the analysis.

You:

I will. Thanks for your time, Yerlan.

Yerlan:

No problem. Talk to you soon!

Vocabulary

evaluation metrics

Meaning: 評価指標

Example: We need to choose evaluation metrics for the new model.

accuracy

Meaning: 正確さ、精度

Example: Accuracy alone isn't enough this time.

latency

Meaning: 遅延、応答速度

Example: We track accuracy, latency, and cost per request.

dataset

Meaning: データセット

Example: Where can we find a good test dataset?

compare results

Meaning: 結果を比較する

Example: Let's compare results on Friday.

cost per request

Meaning: 1リクエストあたりのコスト

Example: We should track cost per request too.

share

Meaning: 共有する

Example: Can you share the dataset with me today?

report

Meaning: 報告書、レポート

Example: I'll also write a short report for the team.

Quiz

1. What does Yerlan think isn't enough this time?

2. What three things does Yerlan suggest tracking?

3. What dataset will they use for testing?

4. When will You and Yerlan compare results?

5. What will You write for the team?

← Back to conversations