AIF-C01 ยท Training and fine-tuning foundation models
Curate, label, size, govern, and represent your data
Data curation, governance, size, labeling, representativeness and reinforcement learning from human feedback.
Key points
- Curate: remove duplicates, errors and noise.
- Label: accurate prompt/response pairs for supervised tuning.
- Size: a few hundred to thousands of good examples, within the model's limits.
- Govern: ownership, privacy, compliance.
- Represent: the real mix of cases and users, to avoid bias.
- Amazon Bedrock: JSONL files in Amazon S3.
- RLHF: human preference rankings align outputs with what people prefer.
Source: https://docs.aws.amazon.com/bedrock/latest/userguide/model-customization-prepare.html
Practice questions
In what file format do you provide training data for an Amazon Bedrock fine-tuning job?
- Excel workbook
- JSONL, one JSON object per line
- A single large PDF
- Parquet only
Show answer
Answer: B. JSONL, one JSON object per line
Training and validation datasets are .jsonl files, one JSON record per line, stored in Amazon S3.
Which actions belong to good fine-tuning data preparation?
- Remove duplicate and low-quality records
- Verify that you have rights to use the data and handle personal data properly
- Include every raw log line without review
- Use only the easiest examples
- Ignore the dataset size limits of the chosen model
Show answer
Answer: A. Remove duplicate and low-quality records, B. Verify that you have rights to use the data and handle personal data properly
Curation removes noise and duplicates; governance checks you may legally use the data. Adding unreviewed data or customer PII without checks harms quality and compliance.
True or false: for fine-tuning, a larger dataset always beats a smaller, carefully curated one.
Show answer
Answer: False
False. Noisy or mislabeled data degrades results; a smaller, clean, representative set often outperforms a larger messy one.
Practice 8 questions on this topic in Bitzu
Short lessons, instant explanations and spaced review of what you get wrong. Free on iOS and Android.
