9.2 Dataset Quality vs Quantity

Post-training data is a control surface, not a bag of examples. A smaller curated set can outperform a larger noisy set in some regimes, as LIMA and AlpaGasus illustrate [1] [2]. That is not a universal inverted scaling law. Results depend on base-model capability, task coverage, diversity, label process, optimization, and evaluation. Quantity still matters when it adds missing support rather than repetitions of the same pattern.

Define Quality as Measurable Properties

“High quality” should expand into testable dimensions:

• correctness: factual, executable, or rubric-valid output;
• relevance: the answer satisfies the request without evasion;
• coverage: task, domain, language, safety, difficulty, and length slices match the product;
• consistency: role format, tool schema, citation policy, and refusal style are coherent;
• diversity: paraphrases do not crowd out rare but important behaviors;
• provenance: license, consent, PII/secrets handling, and deletion obligations are known.

Do not infer that noise always requires exponential compute or that one duplication percentage is optimal. Repetition acts like sample weighting; its effect varies with model, schedule, example difficulty, and evaluation distribution. Measure unique and repeated-token exposure and compare controlled mixtures.

The visualizer is an intuition aid, not an empirical law or a source of default thresholds.

Immutable Manifest and Evaluation Quarantine

Every release candidate should point to an immutable manifest containing source/revision, license or consent basis, collection time, language/domain labels, filter versions, quality scores, mixture weights, tokenizer/template version, shard hashes, and stable sample IDs. Store PII/secrets decisions and deletion lineage so a source can be traced and removed.

Create an evaluation quarantine before curation. It includes public benchmark items, private release prompts, rubrics, semantic neighbors, prior judge examples, and teacher-generated variants. Compare exact hashes, normalized text, and semantic neighbors against every human and synthetic training source. Contamination checks happen before training and again on the final rendered/tokenized corpus.

Deduplicate, Cluster, Then Split

The safe order is:

1. normalize conservatively while preserving raw provenance;
2. exact-deduplicate and detect near duplicates;
3. build semantic clusters for related documents, conversations, entities, and synthetic siblings;
4. assign whole clusters to train/dev/test;
5. compute mixture weights and sample only from the training partition.

Splitting first leaks paraphrases and siblings into validation. Record cluster IDs and split decisions. For preference data, keep both answers and all annotations for a prompt in the same split.

Judges Are Instruments, Not Ground Truth

A reward model logit such as 1.5 is not a portable quality threshold. Scores depend on the reward model’s own chat format, domain, length distribution, and checkpoint. A judge may reward verbosity, familiar phrasing, or its own model family.

Calibrate each judge on a human-labeled gold set from the target slices:

1. use the judge’s documented tokenizer-native format;
2. measure ranking accuracy and calibration by domain, language, length, and safety slice;
3. choose accept/review/reject thresholds from precision/recall and review budget;
4. retain raw scores, rubric version, judge artifact, and rejection reason;
5. send uncertain, disagreement, and high-risk samples to human audit.

Use executable verification for code, math, structured output, and tools when available. Combine independent signals rather than letting one scalar reward decide everything.

Mixture Design and Acceptance Tests

Construct an explicit target distribution. Track per-source examples and tokens, effective sampling probability after filtering, response length, supervised-token fraction, and duplication exposure. Temperature sampling or capped upsampling can protect small important domains, but validate that repeated examples do not dominate updates.

Before a full run, compare candidate manifests with the same checkpoint, tokenizer, token budget, optimization, and evaluation harness. Use multiple seeds when the decision is close. Report paired uncertainty, slice minima, and raw adjudicated examples—not only average win rate.

Accept a data release only if:

• schema, template, EOS, language, and tool-format checks pass;
• license/consent and PII/secrets policy has an owner;
• exact, near, semantic, and benchmark-contamination rates are within declared limits;
• human audit reaches the target precision on critical slices;
• a small SFT probe improves target slices without exceeding base-retention and safety regression limits.

Version the accepted manifest and never mutate it in place. A rejected sample remains useful evidence when its rejection reason is retained.

Quizzes

References

1. Zhou, C., et al. (2023). LIMA: Less Is More for Alignment. arXiv:2305.11206.
2. Chen, L., et al. (2023). AlpaGasus: Training A Better Alpaca with Fewer Data. arXiv:2307.08701.
3. Lee, K., et al. (2021). Deduplicating Training Data Makes Language Models Better. arXiv:2107.06499.
4. Sainz, O., et al. (2023). NLP Evaluation in Trouble: On the Need to Measure LLM Data Contamination for Each Benchmark. arXiv:2310.18018.