TF2-12B: Open 12B Literary Translator

• TF2-12B is a 12B-parameter open model designed for English-to-Romanian literary translation of moral fables using LoRA-based adaptation.
• It utilizes the unified TF2 pipeline to generate high-quality translations with enhanced stylistic fidelity, coherence, and cultural adaptation.
• Deployment optimization through 8-bit quantization ensures near-proprietary performance at a fraction of the cost.

Searching arXiv for the specified TF2-12B paper and related context. TF2-12B is a 12B-parameter open-weight model for English→Romanian literary translation, specifically moral fables, released within the TinyFabulist Translation Framework (TF2). It is a LoRA-adapted version of Gemma-3-12B-it, fine-tuned on TF2’s parallel fable corpus to improve literary translation quality, especially stylistic fidelity, coherence, and cultural adaptation. Within TF2, it functions both as a benchmarked released checkpoint and as a scalable translator for large-scale bilingual dataset generation. The associated study presents TF2 as a unified framework for dataset creation, fine tuning, and evaluation, centered on TF2-12B together with the datasets DS-TF2-EN-RO-15K and DS-TF2-EN-RO-3M (Nadas et al., 9 Sep 2025).

1. Position within the TinyFabulist Translation Framework

TF2 is described as a 4-stage pipeline: Evaluate candidate translators, Create a 15k parallel corpus, Fine-tune open models with LoRA, and Translate the full 3M-story corpus. TF2-12B originates in Stage 3 of this pipeline. The broader framework extends DS-TF1-EN-3M, an English-only synthetic moral-fable corpus, into a bilingual English-Romanian resource. In that sense, TF2-12B is not an isolated model artifact but part of an end-to-end system linking translator selection, supervised adaptation, and large-scale corpus expansion (Nadas et al., 9 Sep 2025).

Within the TF2 model family, TF2-12B is identified as the largest open, self-hostable translation model released for English→Romanian literary translation. The paper positions it as the strongest of the open models in the TF2 family and as the best compromise between quality and deployability among the released checkpoints. A plausible implication is that the model is intended to occupy the operational point where literary adequacy, self-hostability, and deployment efficiency are jointly optimized, rather than maximizing parameter count alone.

The framework assigns TF2-12B two distinct roles. First, after fine-tuning, it is one of the candidate systems used to translate fables in the seed corpus, although GPT-o3 is selected as the reference translator for the silver-standard 15k set. Second, TF2-12B is part of the model family used to translate the much larger DS-TF2-EN-RO-3M corpus. This dual role is central to TF2’s design: the model is both a released research checkpoint and a production instrument for dataset construction.

2. Backbone, adaptation method, and optimization

TF2-12B is based on Gemma-3-12B-it, an instruction-tuned open-weight backbone with 12B parameters. No custom architectural changes to Gemma-3-12B itself are described beyond LoRA adapters and subsequent compression or quantization. Backbone selection followed a benchmarking stage in which multiple proprietary and open systems were evaluated on a held-out fable set; among open backbones, Gemma-3-12B-it was the strongest and was therefore selected for fine-tuning (Nadas et al., 9 Sep 2025).

The fine-tuning objective is framed in instruction-tuning format. The input prompt uses a standardized translation instruction such as “Translate the following fable from English to Romanian:”. The English source is tokenized but masked out of the loss, and only the Romanian target tokens contribute to training loss. The optimization target is therefore the conditional generator

$p(y \mid x),$

where $x$ is the English fable and $y$ is the Romanian translation.

The adaptation mechanism is Low-Rank Adaptation. LoRA is applied to all major projection matrices in the transformer: the self-attention projections q_proj, k_proj, v_proj, o_proj, and the MLP projections gate_proj, up_proj, down_proj. For each weight matrix $W$, the learned update is

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

with $r = 32$, $\alpha = 32$, effective scale $= \alpha/r = 1.0$, and dropout $p = 0.05$. Only the LoRA matrices $A,B$ are trained; the base Gemma-3-12B weights remain frozen during adaptation.

Training uses AdamW, gradient accumulation, cosine learning-rate scheduling, FP16 or bfloat16 mixed precision, and early stopping on validation loss. After convergence, the LoRA adapters are merged into the base model for inference. The paper also describes TF2-12B as part of a two-stage adaptation and deployment process: Stage 1 is instruction fine-tuning with LoRA on DS-TF2-EN-RO-15K, and Stage 2 is compression and deployment optimization after training.

3. Data resources and corpus construction

The principal supervised dataset for TF2-12B is DS-TF2-EN-RO-15K, which contains 15,000 English→Romanian fable pairs sampled from DS-TF1-EN-3M. The set is created using the best-performing LLM translator selected in Stage 1, and it is partitioned into 12,000 train, 1,500 validation, and 1,500 test examples. Each record is stored as JSONL with fields including fable, translated_fable, pipeline_stage, source_lang, target_lang, prompt_hash, llm_name, translation_model, and generation_timestamp. The metadata design is explicitly intended to support traceability and reproducibility (Nadas et al., 9 Sep 2025).

The second major resource is DS-TF2-EN-RO-3M, a large-scale bilingual corpus comprising approximately 3 million English→Romanian fable pairs. It is generated by translating the full DS-TF1-EN-3M English corpus using the fine-tuned TF2 models, especially the stronger checkpoints. The paper emphasizes that the corpus is released openly as a literary translation dataset and that it was produced at very low marginal cost using open models on local or cloud hardware.

The relation between TF1 and TF2 is structurally important. TF1 is defined as English-only synthetic moral fables, whereas TF2 is the bilingual English-Romanian translation of those fables. Consequently, DS-TF2-EN-RO-15K functions as a high-quality translated subset from the TF1 pool, and DS-TF2-EN-RO-3M functions as the large translated expansion. This suggests that TF2-12B is embedded in a data-generation strategy where supervised fine-tuning on a smaller high-quality parallel set is used to enable translation at much larger scale.

A significant methodological constraint is that the “gold” Romanian references are generated by GPT-o3 rather than by human translators. The paper explicitly notes that BLEU and judge scores are therefore relative to synthetic references. That limitation affects how TF2-12B’s performance should be interpreted, especially for questions of literary nuance and human-facing translation fidelity.

4. Evaluation protocol and measured translation quality

TF2-12B is evaluated using two complementary procedures: corpus-level BLEU and a five-dimension LLM-based rubric. BLEU is computed on the test set using GPT-o3 translations as the reference. The paper notes that BLEU is a lexical overlap metric and is not ideal for literary text, but it is retained as a consistency check. BLEU scores are normalized to the interval $x$0, so $x$1 BLEU (Nadas et al., 9 Sep 2025).

For TF2-12B, the reported BLEU values are as follows:

Variant	Temperature	BLEU
FP16 TF2-12B	$x$2	0.0926
FP16 TF2-12B	$x$3	0.0647
FP16 TF2-12B	$x$4	0.0784
8-bit TF2-12B	$x$5	0.0746
8-bit TF2-12B	$x$6	0.0644
8-bit TF2-12B	$x$7	0.0548
Gemma-3-12B-it	—	0.0214

The main evaluation uses an LLM judge scoring five dimensions from 1 to 5: Accuracy, Fluency, Coherence, Style, and Cultural/pragmatic adaptation. The aggregate score is defined as

$x$8

The evaluator is prompted in a professional translation-assessment format and instructed to output valid JSON; the output is parsed for aggregation. Scores are averaged over 100 random test fables per model. The paper also repeats scoring with Grok-3-mini as an independent judge to check for evaluator-family bias, and reports that rankings remain stable.

Relative to its backbone, TF2-12B shows a large improvement in rubric score. Gemma-3-12B-it has average 4.43, whereas TF2-12B achieves 4.83 at $x$9, 4.82 at $y$0, and 4.66 at $y$1. The quantized version scores 4.82 at $y$2, 4.82 at $y$3, and 4.64 at $y$4. At $y$5, the breakdown is Accuracy 4.72, Fluency 4.88, Coherence 4.84, Style 4.87, and Cultural adaptation 4.85. These figures support the paper’s characterization of TF2-12B as especially strong on fluency, style, and cultural adaptation.

In comparison with proprietary systems, the strongest reported average rubric scores are GPT-o3 at 4.92, GPT-4.1 at 4.89, Gemini-2.5-Flash at 4.84, and GPT-o3-mini at 4.83. TF2-12B at 4.83 is therefore close to the reported proprietary baselines. Under the cross-family bias check with Grok-3-mini, GPT-o3 scores 4.92, TF2-12B-Q scores 4.90, and TF2-12B scores 4.85. This suggests that the competitiveness claim is not solely an artifact of a single evaluator family.

5. Compression, quantization, and deployment profile

The paper treats deployment efficiency as a formal part of the TF2-12B pipeline rather than as an afterthought. After LoRA training, adapters are merged, the model is exported in 8-bit quantized form, and the paper also mentions W8A8 quantized models produced with llmcompressor. Deployment artifacts include Safetensors, GGUF for llama.cpp compatibility, and formats usable with vLLM. The resulting efficiency stack is summarized as

$y$6

The paper notes that the quantized TF2-12B is “nearly indistinguishable” from the FP16 version in quality (Nadas et al., 9 Sep 2025).

The quantitative evidence for that claim appears in the rubric scores. TF2-12B FP16 scores 4.83, whereas TF2-12B quantized scores 4.82. The paper presents this as evidence that aggressive quantization has negligible quality loss and is suitable for deployment. The same pattern is reflected in the low-temperature regime more broadly: performance at $y$7 and $y$8 is nearly unchanged, while $y$9 produces a clearer drop. The reported conclusion is that greedy decoding or very low temperature is optimal for translation quality in this setting.

Training and inference hardware are reported explicitly. TF2-12B training was run on 1 × H100 with about 2 hours wall-clock fine-tuning time. Inference and large-scale translation for the 3M corpus were done on 8 × H100 clusters. The released deployment formats are intended to support local inference, vLLM serving, llama.cpp / GGUF workflows, and 8-bit quantized deployments. A plausible implication is that the model is designed to be portable across both research and production environments, including offline settings and commodity-GPU workflows where full-precision serving of a 12B model may be impractical.

6. Cost profile, limitations, and research significance

The paper emphasizes TF2-12B’s role in cost-efficient literary translation. For translating the full 3M-fable corpus, the reported cost estimates are about $W$02,700 for GPT-4.1-mini, about $W$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, and about $350 for the TF2 open model pipeline. The study therefore reports a 97–99% cost reduction versus proprietary APIs at comparable quality (Nadas et al., 9 Sep 2025).

Several limitations are stated directly. First, the reference translations are synthetic rather than human-authored. Second, evaluation uses an LLM judge, which may introduce systematic bias, although the paper partially addresses this with a cross-family judge check. Third, the domain is narrow: fables are structured and formulaic, so results may not generalize to harder literary genres such as poetry or highly idiomatic novels. Fourth, only 100 test samples per model were judged by the LLM evaluator. Fifth, both results and cost comparisons depend on the exact model versions and API pricing available at the time.

These limitations are important for interpreting the broader significance of TF2-12B. The paper’s substantive claim is not that open models have solved literary translation in general, but that a 12B open-weight model, if carefully adapted, can get very close to proprietary translation quality for low-resource literary translation. In that framing, TF2-12B is evidence for a specific methodological recipe: synthetic data, LoRA-based specialization, and quantized deployment can yield a competitive, self-hostable system for a culturally specific translation domain.

The paper also suggests practical research uses for TF2-12B and the TF2 corpora: Romanian literary translation, cross-lingual narrative generation, training a future Romanian fable generator, studying moral reasoning across languages, benchmarking cost-efficient open LLMs, on-device or local deployment for educational or cultural content, and building bilingual datasets for other low-resource literary genres. This suggests that TF2-12B should be understood not only as a translation model but also as a reusable infrastructure component for low-resource literary NLP.