Research

TinyFabulist is my research program on small language models and synthetic stories. The work follows a complete path: specify an English story, translate it into Romanian, then train a compact model to generate Romanian text. Each stage leaves a different question to test, from whether a story follows its instructions to whether an evaluator recognizes a good translation.

The map below separates the published artifacts from the arguments I am developing in the thesis. You can follow the papers for the full methods, or start with the linked notes for one experiment at a time. The recurring question is practical: which property did we actually measure, and what can that measurement support?

TinyFabulist

TinyFabulist is a three-stage research program built around controlled synthetic narratives.

TF1: generation

Structured six-slot specifications are expanded into prompts and passed to open-weight generators no larger than 8B parameters. The released artifact is a three-million-story English corpus with generation metadata and evaluation signals.

• Paper: TF1-EN-3M
• The paper records the dataset schema, generation pipeline, and public release.

TF2: English–Romanian translation

TF2 turns selected TF1 stories into English–Romanian parallel resources and tests how far a fine-tuned open model can narrow the gap to much larger systems at lower inference cost. The project separates the 15K reference set, the three-million-pair corpus, and the released model; they serve different purposes.

• Journal article: Building Large-Scale English–Romanian Literary Translation Resources with Open Models
• Versioned preprint
• The article distinguishes the full three-million-pair corpus from the curated 15K tuning and evaluation set.

TF3: compact Romanian models

TF3 covers tokenizer construction, from-scratch pre-training, compression, evaluation, and Romanian-native generation. The current paper distinguishes the 51.65M-parameter model from the 26.45M-parameter compressed student; neither number should be reduced to “a 50M model” without context.

• Paper: TF3-RO-50M

Evaluation and Romanian NLP

Two supporting lines of work cut across the pipeline:

• a survey of synthetic text and code generation (arXiv);
• Romanian diacritic restoration, including the InnoComp 2025 study published in Springer CCIS 2794 (paper, preprint).

The open-weight panel now has one measured use: under the fixed TF1 protocol, its system ranking tracked o4-mini at Spearman \(\rho=0.93\) and Kendall \(\tau=0.78\). Item-level agreement remained weak, and the planned human arbitration has not run. I therefore use the panel for aggregate comparison under the tested protocol, not as a replacement for human review or as an automatic filter for individual texts.

From the thesis draft

The current proposal-stage thesis joins the three systems under a measurement question: when a pipeline is called controlled, which output properties were actually tested? This series develops the results that are not already covered by the release notes.

• Three Million Stories Are Still a Sparse Sample: What a three-million-row corpus covers when six slots each have one hundred possible values.

• Change One Slot, Watch What Else Moves: A paired intervention test separates response to a requested field from leakage into a field that should stay fixed.

• The Interval Belongs to the Comparison: Why two model gaps require different bootstrap designs—and why one missing record prevents an interval altogether.

• The Panel Was Weak on Items and Useful for Ranking Systems: An open-weight judge panel was weak at item-level agreement yet useful for ranking systems in one fixed protocol.

• When Noisy Text Changes the Winning Model: Romanian diacritic restoration changes winners when clean benchmark text gives way to typos and OCR-like corruption.

• Ten Thousand Adaptation Steps Could Not Remove One Scraper Artifact: One Romanian checkpoint kept emitting a web-page token after supervised adaptation, turning contamination into a model-selection failure.

Reproducibility

Public papers, datasets, prompts, and code improve auditability. They do not make hardware behavior, third-party model serving, or stochastic execution automatically identical. Each technical note records the boundary it actually tested.