Handling Evaluation is usually about compatibility, sharing, archiving or saving space; define the goal before choosing an approach. Token budgets affect eval size/cost; sample first.
Why it matters
Different goals imply different formats, tools and quality settings.
Long eval sets cost more.
Step by step
- Confirm the model context and output budget.
- Paste the prompt/document and estimate tokens/usage.
- Above 80-90%, split, summarize or retrieve.
- Count system prompts/tool schemas too.
- Verify billing/hard limits with the official tokenizer.
Common mistakes
- Treating character count as token count.
- Ignoring output/system prompt usage.
- Assuming all models use the same tokenizer.
- Stuffing a long document instead of retrieving/summarizing.
Checklist
- Model context confirmed.
- Output budget reserved.
- System/tool prompts counted.
- Long text retrieved/split.
- Critical cases checked with the official tokenizer.
Key facts
- Long eval sets cost more.
- Stratified sampling covers cases.
FAQ
Why should I handle Evaluation?
Handling Evaluation is usually about compatibility, sharing, archiving or saving space; define the goal before choosing an approach. Token budgets affect eval size/cost; sample first.
How long does it take?
Usually 1-3 minutes; for intermittent issues, run 30 seconds and retest elsewhere.
Is data uploaded?
Runs locally and does not upload video/audio; extensions may still read the page.
Source: OpenAI evals