The LLM-Synthetic-Data list claims July 2025 currency and was last touched in April 2026

The repository is a licence and a markdown file

There are two entries at the top level, `LICENSE` and `README.md`, and the recorded primary language is empty because there is no code. The project is 492 stars and 40 forks, which is a large audience for a document, and the pitch in the opening paragraph is a claim about maintenance and classification rather than about content: it collects the most live-updated and finely categorized work on synthetic data for language models, covering papers, tools, datasets and blogs. The file then asks readers to follow and star it, and marks entries it highly recommends with a flame glyph, a convention that is worth understanding before reading the list, since an unmarked entry is not a criticism and a marked one is not a citation. Two other curated lists on the same subject are linked, both marked as top picks, which is a useful signal about the ecosystem: the reading list is partly a directory of reading lists.

The currency claim is in the title, the evidence is structural

The document's headline and the repository description both say it is updated to July 2025, and the last commit on the default branch main is dated 9 April 2026. So the label has not moved in nine months of repository activity, and a reader has no way to tell from the file whether that is because the field is quiet or because the file is not being maintained. The section headed latest updates is the place to look, and it describes four changes with no dates attached: domain-specific synthesis surveys were added to the surveys section, the methods section was reorganized by training stage with what it calls ultra-fine subcategories and is marked highly recommended, a new section was created for analyses of synthetic data, and the application section was expanded to nineteen sub-areas. Every one of those is an editorial act on the structure of the list rather than an addition of recent work. For a bibliography whose argument is that it tracks a fast-moving field, that is the gap to hold in mind.

Four of the table-of-contents anchors are misspelled

The contents list is written by hand rather than generated, and four of its anchors do not match the headings they point at. The refinement learning entry links to `#45-refment-learning`, with the second i missing. The synthetic and real data entry links to `#47-synthetci-and-real`, which is both misspelled and a different fragment from the heading it labels. The effect of synthetic data entry links to `#51-effect-syntectic-data`, with the h transposed. The writing entry links to `#67-wirting`. GitHub derives a heading's fragment from its own text, so these four jump to nothing while their neighbours work, which is the worst failure mode for a navigation list: the link looks right and goes nowhere. The fix is mechanical, and the fact that it is not done is a reasonable measure of how much attention the hand-maintained parts of the file receive.

The methods section is sorted by training stage, which is the useful part

The one structural decision here worth borrowing is how the method literature is cut. Rather than sorting by technique, the section is organised by where in training the method applies, with subcategories for pre-training, continue pre-training, instruction tuning, alignment, refinement learning, LLM benchmarking, and using synthetic and real data jointly. That axis answers the question a practitioner actually has, which is when in the pipeline a given technique is available and what it competes with. Refinement learning as its own stage is a category most lists fold into post-training, and having synthetic and real data treated as a joint problem rather than as synthetic versus real is the framing that keeps a list honest, since almost nobody ships pure synthetic data. The file marks this section as highly recommended, and the 4.1 entry visible in the pre-training subsection is the Phi-4 technical report from Microsoft Research.

Nineteen application sub-areas, from mathematics to federated learning

The application section runs to nineteen numbered sub-areas, which is the fine-grained claim in action. They start at mathematical reasoning, code generation, agent and tool use, vision and language, retrieval-augmented generation and long context, then move through writing, AI for science, text-to-SQL, the synergy between large and small models, weak-to-strong generalisation, distilling a small model, multilingual data, structured data, natural language understanding, logic reasoning, dialogue systems, federated learning, and generative design. Read as a list of headings it is unremarkable; read as a claim about how the field is subdividing, it is more informative than most. Two of the entries are notable for being about the relationship between models rather than about an application, which is the same instinct that produces the joint synthetic and real category in the methods section. A blog entry and seven blog posts are dated and attributed individually, so the sections are not uniformly treated.

The blog section is dated and attributed, the surveys section names venues

The entries are not annotated consistently, and the difference tells you how much to trust each. The blog section gives a date and an author for all seven: a Hugging Face post on saving cost, time and carbon, two entries from the OpenAI cookbook, a self-instruct piece, a Google research post on CodeLM, an agentic data generation post, a Medium article on generation, curation and evaluation, and a guide from Confident AI, running from February 2024 to November 2024. The surveys section does something more useful: it names a venue for each, so a COLM 2024 best-practices survey, an ACL Findings 2024 survey, an EMNLP 2024 survey on data annotation, several arXiv postings from 2024 and 2025, and one OpenReview entry from 2025. Where a venue is given you can check whether it was peer reviewed. Where it says only arXiv, you cannot, and the file does not flag the difference.

Editorial conclusion

Use this list as an index rather than as a survey, because that is what it is, and its real value is the taxonomy rather than the coverage. Splitting the method literature by training stage, from pre-training through continue pre-training, instruction tuning, alignment, refinement learning and benchmarking, and then adding a section on using synthetic and real data jointly, is an axis most bibliographies of this literature do not use, and it is the part worth borrowing. Treat the freshness claim with care. The title and the description both say updated to July 2025, the last commit on the default branch main is dated 9 April 2026, and the change log under the heading for latest updates describes four structural edits without dates, so nothing in the repository tells you which entries are current and which are two years old. Check the venues before you cite anything here, since several entries are preprints and the file does not distinguish a published survey from an arXiv posting except in the venue it names. The last commit date is 9 April 2026 and there are no releases.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. pengr/LLM-Synthetic-Data on GitHub
4. README