DS-TF2-EN-RO-3M Fable Translation Corpus

• DS-TF2-EN-RO-3M is a large-scale synthetic English–Romanian corpus of moral fables that bridges DS-TF1-EN-3M to bilingual literary translation.
• It employs a four-stage pipeline, starting with high-quality GPT-o3 translations and followed by fine-tuning Gemma-based models for mass translation.
• The resource underpins cost-efficient NLP research, low-resource literary translation, and Romanian language modeling with narrative nuance.

Searching arXiv for the cited TF2 and TF3 papers to ground the article. Search results for arXiv query ([2509.07829](/papers/2509.07829)) OR ([2601.10410](/papers/2601.10410)) DS-TF2-EN-RO-3M [TinyFabulist Translation Framework](https://www.emergentmind.com/topics/tinyfabulist-translation-framework-tf2) DS-TF2-EN-RO-3M is a large-scale synthetic English–Romanian parallel corpus for literary machine translation, specifically centered on short moral fables. Within the TinyFabulist Translation Framework (TF2), it functions as the bilingual extension of DS-TF1-EN-3M, transforming a corpus of roughly 3 million synthetic English fables into an English–Romanian resource for low-resource literary translation, cross-lingual narrative generation, benchmark development, and related research on cost-efficient open-model NLP in culturally meaningful domains (Nadas et al., 9 Sep 2025). In the TF1–TF2–TF3 progression, TF2 serves as the translation-based bilingual bridge between the original English corpus and later Romanian-centric modeling pipelines, with TF3 explicitly using the Romanian split of TF2 as a foundation for tokenizer training and Romanian language-model pretraining (Nadas et al., 15 Jan 2026).

1. Position within the TinyFabulist lineage

TF2 extends the earlier TinyFabulist Framework / TF1 from a monolingual English story-generation resource into a bilingual translation setting. TF1 introduced DS-TF1-EN-3M, described as the largest collection of synthetic English fables to date, and TF2 converts that English source pool into a parallel English–Romanian literary machine translation resource (Nadas et al., 9 Sep 2025).

The dataset name DS-TF2-EN-RO-3M denotes the main large-scale bilingual corpus released in TF2. The paper presents it alongside DS-TF2-EN-RO-15K, a smaller “silver-standard” parallel set used as the key supervision source and benchmark reference set. This establishes a two-tier resource structure: a smaller benchmark-quality seed corpus and a much larger expanded corpus generated by adapted open models (Nadas et al., 9 Sep 2025).

In TF3-RO, TF2 is treated as the intermediate stage between the original English TinyFabulist corpus and a Romanian-native generation pipeline. TF3 describes TF2 as having “extended the dataset with high-quality English--Romanian translations and released the first large-scale bilingual synthetic corpus of moral fables via Hugging Face (ds-tf2-en-ro-3m)” (Nadas et al., 15 Jan 2026). This positions DS-TF2-EN-RO-3M not merely as an auxiliary artifact, but as the bilingual bridge resource from which later Romanian-only corpora can be extracted.

2. Corpus composition and data model

DS-TF2-EN-RO-3M is a parallel English–Romanian corpus whose source side is the English fable collection from DS-TF1-EN-3M and whose target side is machine-generated Romanian. It is synthetic on both sides: the English texts originate from the prior TF1 generation pipeline, and the Romanian targets are produced by LLMs rather than human translators (Nadas et al., 9 Sep 2025).

The operational translation direction in the experiments is English $\rightarrow$ Romanian. The genre is tightly controlled: the dataset consists of short moral fables, typically concise narrative texts ending in an explicit moral. The paper treats this genre control as central because literary translation quality in this setting depends not only on semantic accuracy but also on narrative flow, style, and cultural or pragmatic adaptation (Nadas et al., 9 Sep 2025).

The TF3 paper reinforces this characterization by stating that the Romanian split of TF2 “consists of synthetic moral fables translated from the original English TinyFabulist dataset under strict structural constraints” (Nadas et al., 15 Jan 2026). It also supports the interpretation that the dataset is story-level aligned, since TF2 is described as extending the original English fable dataset with Romanian translations rather than as a sentence-shuffled or document-misaligned collection.

The paper provides limited descriptive statistics for the full 3M corpus. It specifies that stories average about 450 tokens per language, so an English–Romanian pair is about 900 tokens on average. Elsewhere, it states that each input sequence contains on average 300–400 tokens and outputs average 350–450 tokens, while the cost section estimates roughly 300–600 input tokens and 300–600 output tokens per fable. On that basis, the full translation of 3M fables is estimated at 1.8–3.6 billion total tokens, with a mid-case of 2.7B tokens assuming 450 input and 450 output tokens per example (Nadas et al., 9 Sep 2025).

The paper does not provide token totals, vocabulary size, deduplicated corpus size after filtering, or quality-tier breakdowns for DS-TF2-EN-RO-3M. It also does not define train/dev/test splits for the 3M corpus itself; only the 15K set has explicit splits (Nadas et al., 9 Sep 2025).

3. Construction pipeline and relation to DS-TF2-EN-RO-15K

The relation between DS-TF2-EN-RO-15K and DS-TF2-EN-RO-3M is central to understanding the dataset. The 15K set is created first by sampling 15,000 English fables from DS-TF1-EN-3M and translating them with the best-performing Stage 1 translator, identified as GPT-o3 based on rubric evaluation. This 15K corpus is described as a “silver-standard” parallel set and is split into 12,000 training, 1,500 validation, and 1,500 test examples (Nadas et al., 9 Sep 2025).

The 15K set is stored in JSONL format and includes the fields fable, translated_fable, pipeline_stage, source_lang, target_lang, prompt_hash, llm_name, translation_model, and generation_timestamp. The prompt_hash is explicitly described as a SHA-256 hash used for deduplication and reproducibility. The paper does not detail additional post-translation filtering, confidence scoring, or human revision for this 15K set; quality assurance comes mainly from selecting the strongest translator through Stage 1 benchmarking (Nadas et al., 9 Sep 2025).

TF2 describes a four-stage workflow. In Stage 1, 13 candidate systems, including proprietary LLMs, commercial MT services, and open models, are benchmarked using an LLM-as-a-judge rubric rather than BLEU for initial model selection, because literary translation quality is not well captured by lexical-overlap metrics. In Stage 2, the 15,000 short fables are translated with GPT-o3 to produce DS-TF2-EN-RO-15K. In Stage 3, open Gemma backbones are fine-tuned on that 15K parallel corpus. In Stage 4, the best fine-tuned model is used for mass translation of the entire TF1 corpus, producing DS-TF2-EN-RO-3M (Nadas et al., 9 Sep 2025).

This distinction is methodologically important. DS-TF2-EN-RO-15K is the direct supervised fine-tuning dataset, whereas DS-TF2-EN-RO-3M is the resulting large-scale translated corpus and released resource. The paper does not say that the full 3M parallel set was itself used to train TF2-12B or the other TF2 models. Instead, model development proceeds from the 15K set, and the 3M corpus appears as the product of that pipeline rather than as its direct supervision source (Nadas et al., 9 Sep 2025).

The paper further states that the full English corpus of approximately 3 million AI-generated English fables was translated “using our best-performing system, as well as other fine-tuned checkpoints where appropriate,” but it does not break down which subset, if any, was translated by which checkpoint. A plausible implication is that per-example provenance is expected to be recoverable from metadata, since the schema includes llm_name and translation_model (Nadas et al., 9 Sep 2025).

4. Fine-tuned open models underlying the corpus

DS-TF2-EN-RO-3M is not translated directly with GPT-o3. Instead, GPT-o3 is used to create the high-quality 15K seed/reference corpus, after which fine-tuned open Gemma-based models generate the large-scale expansion (Nadas et al., 9 Sep 2025). The paper says the process follows instruction-tuning practices, although its concrete formulation is standard supervised translation fine-tuning in instruction format.

Inputs are converted into prompt-response pairs such as “Translate the following fable from English to Romanian:” followed by the source text. Tokenization uses each model family’s native SentencePiece or BPE vocabulary. In the labels, only Romanian target tokens contribute to the loss; English prompt and source tokens are masked out (Nadas et al., 9 Sep 2025).

The LoRA update is given as

$\Delta W=\tfrac{\alpha}{r}BA$

with rank $r=32$, $\alpha=32$, effective scaling 1.0, and dropout $p=0.05$. Adapters are inserted into q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, and down_proj. Training uses AdamW, gradient accumulation, cosine learning-rate scheduling, and mixed precision (FP16 or bfloat16), with early stopping on validation loss. The paper does not specify batch size, learning rate, number of epochs, warmup, or maximum sequence length (Nadas et al., 9 Sep 2025).

After convergence, LoRA adapters are merged into the backbone. The authors also produce 8-bit quantized variants and W8A8 quantized models via llmcompressor, with exports including GGUF and compatibility with vLLM and llama.cpp. The abstract describes the process as two-stage fine-tuning: instruction tuning to capture genre-specific narrative style, followed by adapter compression for efficient deployment (Nadas et al., 9 Sep 2025).

The flagship open model is TF2-12B. Relative to its untuned backbone Gemma-3-12B-it, it improves from an average rubric score of 4.43 to 4.83 at $T=0.0$, with BLEU rising from 0.0214 to 0.0926. The corresponding TF2-12B-quant model is nearly identical in quality, reaching 4.82 average rubric score at $T=0.0$. The strongest proprietary system, GPT-o3, scores 4.92, while a cross-family judge-bias check using Grok-3-mini as evaluator reports 4.90 for TF2-12B-Q versus 4.92 for GPT-o3 reference translations (Nadas et al., 9 Sep 2025).

Temperature ablations affect the released 3M corpus directly. The paper finds that low temperatures, especially $T=0.0$, consistently yield the strongest literary translation quality and states that $T=0.0$ is adopted as the default for “all subsequent large-scale translation runs.” This indicates that DS-TF2-EN-RO-3M is intended to reflect deterministic, quality-maximizing decoding rather than diversity-oriented sampling (Nadas et al., 9 Sep 2025).

5. Evaluation logic and research uses

Although DS-TF2-EN-RO-3M is the principal large-scale release, it is not the main benchmark in the reported experiments. The held-out split from DS-TF2-EN-RO-15K serves that role because it contains the GPT-o3 “silver-standard” references needed for direct comparison. BLEU is computed against those Romanian references, and the literary rubric is also applied on held-out examples (Nadas et al., 9 Sep 2025).

The evaluation framework combines corpus-level BLEU with a five-dimension LLM rubric: Accuracy, Fluency, Coherence, Style, and Cultural/Pragmatic Adaptation, each scored from 1 to 5. The average rubric aggregation is given explicitly as

$$\text{Avg.\ Score}=\frac{\text{Accuracy}+\text{Fluency}+\text{Coherence}+\text{Style}+\text{Cultural/Pragmatic}}{5}.$$

The Stage 1 benchmark uses this rubric on 100 randomly selected test instances per model, with the evaluation prompt instructing the judge to act as a professional translation evaluator and to output JSON scores and justifications (Nadas et al., 9 Sep 2025).

The dataset’s intended uses extend beyond evaluation. The paper positions DS-TF2-EN-RO-3M as a resource for low-resource English–Romanian machine translation, especially creative or literary translation where style, coherence, and cultural adaptation matter more than in news MT. It is also framed as useful for fine-tuning, benchmark development, cross-lingual narrative generation, cross-lingual moral reasoning, and research on cost-efficient open-model NLP in culturally meaningful domains (Nadas et al., 9 Sep 2025).

TF3-RO demonstrates one such downstream use. Rather than training on the English side, TF3 extracts and uses the Romanian split of TF2 for tokenizer learning, pretraining corpus construction, distillation, evaluation design, and linguistic justification (Nadas et al., 15 Jan 2026). The Romanian TF2 corpus is packed into 2,048-token blocks for causal language-model training, and the total TF3 training corpus derived from Romanian TF2 contains approximately one billion tokens, depending on tokenizer choice (Nadas et al., 15 Jan 2026). This shows that DS-TF2-EN-RO-3M supports not only translation research but also monolingual Romanian LM development when one side of the bilingual resource is isolated.

6. Tokenization, morphology, and Romanian-specific significance

A major significance of DS-TF2-EN-RO-3M lies in what it reveals about Romanian as a morphologically rich and computationally under-resourced language. TF3 states that TF2 showed Romanian translations required roughly 85% more tokens under a standard multilingual vocabulary than the English originals, despite only modest character-length increase. Elsewhere, this is paraphrased as requiring nearly twice as many tokens under standard multilingual tokenizers (Nadas et al., 15 Jan 2026).

The paper connects this token inflation to several technical effects: reduced effective context length, higher compute cost, poorer segmentation of Romanian stems and affixes, and less efficient inference and deployment. Diacritics are treated as important signals for segmentation and morphology, so TF3 preserves them in the Romanian TF2 text and applies only light-weight normalization: removing trailing whitespace, filtering empty examples, standardizing line breaks, and avoiding both diacritic stripping and aggressive cleaning (Nadas et al., 15 Jan 2026).

TF3 uses the Romanian split of TF2 to build Romanian-specific SentencePiece tokenizers, both BPE and Unigram, with an approximately 32k-vocabulary. In tokenization experiments over the Romanian training corpus derived from TF2, average sentence lengths before packing are reported as follows (Nadas et al., 15 Jan 2026):

Tokenizer	Average tokens/sentence	Range
Romanian BPE	304.89	227–887
Romanian Unigram	340.35	243–902

These statistics are not presented as TF2 release metadata in the narrow sense, but they are directly tied to the Romanian side of TF2 as consumed by TF3. Their broader significance is that DS-TF2-EN-RO-3M provides empirical evidence for tokenizer mismatch in Romanian under multilingual vocabularies, thereby motivating Romanian-specific tokenization and subsequent Romanian-native modeling (Nadas et al., 15 Jan 2026).

The TF3 paper also notes an internal schema tension. TF1 is described as having a six-slot scaffold, “character -> trait -> setting -> conflict -> resolution -> moral,” whereas preprocessing of translated Romanian TF2 text is described as following a fixed scaffold of five components: character, setting, challenge, resolution, and moral (Nadas et al., 15 Jan 2026). The paper itself does not resolve this discrepancy. A plausible implication is that the preprocessing description abstracts away one field rather than denying the original six-slot generation process, but that remains interpretive rather than explicitly stated.

7. Cost profile, openness, and limitations

The paper’s cost argument is central to the value proposition of DS-TF2-EN-RO-3M. Translating the full 3M-fable corpus with proprietary APIs is estimated to cost approximately $\Delta W=\tfrac{\alpha}{r}BA$013,500 for GPT-4.1, $\Delta W=\tfrac{\alpha}{r}BA$1270,000 for DeepL <a href="https://www.emergentmind.com/topics/adversarial-prompt-injection-api" title="" rel="nofollow" data-turbo="false" class="assistant-link" x-data x-tooltip.raw="">API</a> Pro. By contrast, generating the full 3M translations with the open TF2 setup is estimated at about $350 in compute, which the authors interpret as a 97–99% cost reduction while preserving near-parity translation quality on the rubric (Nadas et al., 9 Sep 2025).

The appendix gives the cost formula as

$\Delta W=\tfrac{\alpha}{r}BA$2

where $\Delta W=\tfrac{\alpha}{r}BA$3 and $\Delta W=\tfrac{\alpha}{r}BA$4 are input and output prices per million tokens and $\Delta W=\tfrac{\alpha}{r}BA$5 denotes hidden reasoning tokens for reasoning models (Nadas et al., 9 Sep 2025).

The release is explicitly described as open, under MIT licensing, with datasets on Hugging Face and accompanying scripts, prompts, evaluation prompts, code, and model weights. Deployment artifacts include GGUF, Safetensors, quantized checkpoints, and compatibility with vLLM and llama.cpp (Nadas et al., 9 Sep 2025). A minor reproducibility caveat is that the manuscript gives two slightly different Hugging Face naming variants for the 3M dataset: klusai/tf2-en-ro-3m in one place and klusai/ds-tf2-en-ro-3m in another, even though the dataset name in the text is consistently DS-TF2-EN-RO-3M (Nadas et al., 9 Sep 2025).

The limitations are substantial and directly affect interpretation. First, the dataset is synthetic on both sides, and the Romanian references are synthetic rather than human translations; the paper explicitly notes that these “ground truth” references may contain subtle errors or unnatural phrasing. Second, the domain is narrow: moral fables are short, relatively formulaic, and end with explicit morals, which the authors caution may represent an upper bound for low-resource literary MT performance rather than a result that generalizes to novels, poetry, or more stylistically complex literature. Third, because the Romanian side of DS-TF2-EN-RO-3M is generated by fine-tuned open models, it may contain translation artifacts, literal renderings, awkward word choices, and inherited biases from the generation pipeline (Nadas et al., 9 Sep 2025).

Additional limitations concern evaluation and linguistic transfer. The framework relies heavily on LLM-as-a-judge evaluation, which may introduce stylistic or family bias, though the paper partially mitigates this with a Grok-3-mini cross-check. More broadly, the challenge of correctly adapting moral and pragmatic content for Romanian audiences remains open, which is why Cultural/Pragmatic Adaptation is treated as a distinct rubric dimension (Nadas et al., 9 Sep 2025). TF3 extends this perspective by treating translation-based Romanian text as conceptually distinct from Romanian-native generation, thereby motivating the later move from TF2’s bilingual translated corpus to TF3’s Romanian-native synthetic corpus (Nadas et al., 15 Jan 2026).

In the paper’s own framing, DS-TF2-EN-RO-3M matters because it is the first openly released large-scale English–Romanian parallel literary corpus of moral fables, built through a reproducible and low-cost pipeline. It links TF1’s 3M English fables to TF2’s fine-tuned open translation models and, through TF3, to Romanian-native model development, thereby establishing a synthetic-data pathway from English story generation to bilingual literary translation and then to Romanian-centered language modeling (Nadas et al., 9 Sep 2025, Nadas et al., 15 Jan 2026).