The three-million-item Romanian corpus is TF3 output, not TF3 training input. Reversing that arrow is the provenance mistake this map is meant to prevent.

Stage Starts with Produces Strongest reported evidence
TF1 six-slot fable specifications three million English fables ten-model comparison; 100 prompts per model
TF2 TF1 English fables and 15K GPT-o3 Romanian silver references 15K tuning set, three-million-pair corpus, 1B/4B/12B adapters 12B rubric score 4.43 → 4.83; BLEU 0.0214 → 0.0926
TF3 translated Romanian training text 51.65M teacher, 26.45M student, three million direct-Romanian fables teacher cross-entropy ≈0.89 and perplexity ≈2.43; agreement, entity, and grammar probes

Follow one record through the chain

A TF1 record begins with a structured request, becomes an English fable, and keeps its prompt, model, decoding setup, and generation metadata. The TF1 paper and release record describes three million such records and compares ten generators on 100 prompts each.

TF2 takes the English text as source material. GPT-o3 supplies 15,000 Romanian silver references, of which 12,000 train the adapters. The tuned models then support the three-million-pair release. That order explains why the 15K set, scale corpus, and model checkpoints need different names. The TF2 paper records the artifacts, the 4.43-to-4.83 rubric change, and the BLEU comparison.

TF3 consumes translated Romanian training text and starts a 51.65M-parameter model from random weights. Compression yields a separate 26.45M student. Only after training does the pipeline generate three million new fables directly in Romanian. The TF3 paper carries the tokenizer, training, compression, and probe results.

The breakpoints

A good TF1 judge score says little about translation fidelity. A strong TF2 translation score does not validate the TF3 student. Low TF3 validation loss does not certify the generated corpus as natural Romanian. The synthetic-data survey (arXiv) and current judge work exist because every arrow introduces a new measurement problem.

The maintained map lives at /research/; this snapshot preserves the provenance order that readers most often reverse.