Synthetic Data Can Make the Model Worse

TL;DR for operators

A team with authoritative domain documents but little labeled training data has an attractive option: ask a capable model to manufacture question-answer pairs, then fine-tune a smaller open model on them. The operational risk is assuming that domain relevance makes those examples safe training material.

In this paper, a simple synthetic-data pipeline moved LLaMA 3.1 8B backward on open-ended legal QA: its LegalMC4 score fell from 43.0% to 35.4%. A more structured pipeline raised the same score to 55.4%. Across both LLaMA 3.1 8B and Gemma 3 12B, that structured treatment improved all four tested German legal benchmarks.¹

The practical lesson is not that one particular prompting recipe has been proved superior. The successful treatment bundles several changes: different kinds of questions, a separate quality reviewer, and a larger retained dataset. What the evidence supports is a data-production system. For teams specializing open models, the relevant design decision is therefore what supervision enters training, not merely how cheaply synthetic examples can be generated.

Domain-relevant supervision can still train the model in the wrong direction

Synthetic data is often framed as an annotation substitute. If a company owns authoritative policies, manuals, statutes, or technical documentation, generating training examples from those materials appears to solve the label shortage without paying for every example manually.

The paper tests why that logic is incomplete.

Its simpler condition generated 15,734 German legal QA pairs through a one-step instruction process. For Gemma 3 12B, this produced useful gains on open-ended questions: BGB QA rose from 36.8% to 46.6%. But the same treatment barely moved either multiple-choice benchmark, and it damaged LLaMA 3.1 on three of four legal evaluations. LegalMC4 open-ended QA fell 7.6 percentage points; BGB QA fell 1.6 points.

That result matters because the source material was not unrelated web text. The examples were generated from German statutory material. Domain grounding alone did not guarantee useful adaptation.

The replacement question for an operator is therefore more demanding: what kinds of behavior does the synthetic dataset repeatedly teach the model?

The stronger pipeline changes both the questions and the admission standard

The paper’s stronger treatment deliberately varies what the synthetic questions require.

Some examples ask for direct comprehension of statutory clauses. Others paraphrase questions into client-style language, introduce factual scenarios requiring application of a provision, or require reasoning across multiple provisions. This is the paper’s difficulty-graded QA design: synthetic examples differ not only in wording but in the type of legal reasoning demanded.

Generation is followed by a separate quality gate. A reviewer checks whether each candidate is answerable, supported by the statute, and non-redundant. Failed candidates are discarded rather than repaired or replaced.

This filter is substantial. The generator produced 31,777 candidates and retained 23,905:

These counts are best read as an implementation description of the quality-control stage, not as proof that rejection itself caused the benchmark gains. In particular, the experiments do not independently vary difficulty grading and reviewer filtering.

That distinction becomes important when interpreting the main result.

The bundled pipeline improves every tested legal benchmark

Table 2 is the paper’s main comparative evidence. With LoRA configuration and optimization settings held fixed, both model families improve across all four legal benchmarks after training on the difficulty-graded, filtered data.

The largest movement is Gemma’s BGB open-ended result: 36.8% to 76.4%, a 39.6-point increase. LLaMA’s corresponding gain is 20 points.

The LegalMC4 results are also operationally interesting because those questions contain heterogeneous legal material and distracting surrounding context. The paper treats this as an approximation of the challenge facing the generation stage after retrieval: the model has relevant material somewhere in its context but must still identify and use it correctly.

That does not make the experiment an end-to-end RAG evaluation. Retrieval quality, document selection, ranking, and pipeline interactions are not tested.

Specialization did not produce a large general-capability collapse

A reasonable deployment concern is that domain adaptation may improve the target workflow by degrading the model elsewhere.

The paper uses ARC-Easy, ARC-Challenge, a German ARC subset, and MMLU as a sensitivity check on this possibility. The results are broadly stable rather than uniformly unchanged.

For Gemma with the difficulty-graded treatment, ARC-Easy rises 2.1 points, ARC-Challenge 2.9, and ARC-DE 1.9, while MMLU falls 0.6. LLaMA moves between +0.4 and -1.7 points across the four measures.

These benchmarks do not prove that every general capability has been preserved. They do provide evidence against substantial catastrophic forgetting within the tested evaluation set. For an organization that wants one open model to retain broad utility while gaining competence in a recurring specialist workflow, that is materially different from requiring a completely isolated specialist model.

For domain teams, synthetic data becomes a production-control problem

Cognaptus inference: the business decision exposed by these experiments sits upstream of fine-tuning.

For a team holding authoritative source material, the relevant workflow is:

source normalization → controlled example generation → independent admission checks → contamination-aware splitting → fine-tuning → domain and general evaluation

The affected user is the team deciding whether to invest in domain adaptation. The condition is a setting where trustworthy source documents exist but labeled examples are scarce. The decision is how much engineering to allocate to the synthetic-data pipeline before paying for training runs.

The paper gives one indicative compute point: its LoRA fine-tuning runs used an 80 GB NVIDIA A100 and took roughly 9–10 hours per model-dataset run. That makes weak supervision more than a dataset-quality issue. Poor examples consume training infrastructure and can leave the resulting model worse on the task that justified the run.

A separate reviewer is therefore best understood as a training-data control, not merely a cleanup feature.

What the experiment does not separate

The main interpretive boundary is causal.

The successful condition simultaneously changes the distribution of question types and adds reviewer filtering. It also retains 23,905 examples versus 15,734 in the simple condition. Because these factors are not independently ablated, the experiment cannot establish how much of the gain comes from difficulty diversity, filtering, dataset size, or their interaction.

Open-ended answers are also scored by GPT-4.1 rather than by a paper-specific human validation exercise. The evidence covers two open model families and German statutory/legal-document tasks. Generalization to another jurisdiction, internal corporate documents, clinical material, or a different model architecture remains an engineering hypothesis to test rather than a demonstrated result.

The proprietary-model comparisons deserve similar restraint. The adapted Gemma model is competitive with GPT-5-mini on the multiple-choice tasks and scores above the reported GPT-5-family results on BGB open-ended QA, but the authors explicitly note that reasoning behavior cannot be matched cleanly across those systems.

Train on selected supervision, not generated supervision

The paper changes a practical assumption about synthetic domain adaptation.

Authoritative documents can be converted into scalable supervision, and within these German legal experiments the resulting specialization can be substantial without a corresponding collapse on general benchmarks. But automatically generated QA is not valuable simply because it came from the right corpus.

The operational unit that deserves evaluation is the synthetic-data pipeline: what reasoning patterns it generates, what examples it rejects, how contamination is controlled, and whether the resulting model improves on held-out workflows.

For organizations already capable of generating thousands of examples cheaply, that shifts the next optimization target from generation throughput to deciding which generated supervision deserves to alter the model.

Cognaptus: Automate the Present, Incubate the Future.