DenizGogus/denela-translator-models

Denela Translator Models

Denela is an experimental AI-assisted constructed language with its own grammar, lexicon, derivation rules, and orthography. English and Turkish are used as reference meaning channels for training and inspection, not as the identity of the language itself. This repository contains local transformers checkpoints for translating:

• English -> Denela
• Turkish -> Denela
• Denela -> English
• Denela -> Turkish

The hosted demo/API is intended to call these local checkpoints directly. It does not need an external LLM at inference time.

Repository Structure

denela-translator/
to-denela-translator/
from-denela-translator/

to-denela-translator is used for English/Turkish -> Denela.

from-denela-translator is used for Denela -> English/Turkish.

denela-translator is an earlier/general checkpoint kept for compatibility.

Intended Use

This model is intended for:

• translating between English, Turkish, and Denela
• powering the Denela web/API demo
• experimenting with human-learnable synthetic-language translation
• inspecting generated Denela with dictionary/grammar breakdowns in the companion app

This model is not intended for:

• production-grade legal, medical, or safety-critical translation
• general English <-> Turkish translation
• translating arbitrary languages outside English, Turkish, and Denela
• treating Denela as a stable natural language

Usage

from transformers import AutoModelForSeq2SeqLM, AutoTokenizer

repo_id = "DenizGogus/denela-translator-models"

tokenizer = AutoTokenizer.from_pretrained(repo_id, subfolder="to-denela-translator")
model = AutoModelForSeq2SeqLM.from_pretrained(repo_id, subfolder="to-denela-translator")

prompt = "translate English to Denela: I want to see the stars tonight."
inputs = tokenizer(prompt, return_tensors="pt")
outputs = model.generate(
    **inputs,
    max_new_tokens=128,
    num_beams=8,
    early_stopping=True,
)

print(tokenizer.decode(outputs[0], skip_special_tokens=True))

For Denela -> English/Turkish, load from-denela-translator and use prompts like:

translate Denela to English: ka na hira taraai naku
translate Denela to Turkish: ka na hira taraai naku

Example Outputs

Input: translate English to Denela: I want to see the stars tonight.
Output: ka na nevir hira taraai naku

Input: translate Denela to Turkish: ka na hira ka en nara en tara
Output: Arkadaşımın yıldızını görüyorum.

Input: translate Turkish to Denela: Seni ve beni seviyorum.
Output: ka na leva tu ko ka

Training Data

The model was fine-tuned from google/byt5-small on synthetic parallel examples for English, Turkish, and Denela. The training pipeline uses:

• a hand-designed Denela grammar
• an approved Denela lexicon
• canonical Denela form normalization
• synthetic sentence triples
• text-source sentence adaptation for broader coverage

The Denela v0.2 local training build used:

• 43,448 source translation triples before quality filtering
• 456,532 weighted training pairs
• 9,298 unweighted, group-isolated validation pairs
• zero normalized pair, input, triple, or split-group leakage
• relation-focused contrasts for ko/ke and fu/su/mu

v0.2 Evaluation

The 24-case relation/modality suite improved from 16.7% to 100% after clean retraining. On a fixed 250-example sample per direction, the promoted checkpoints achieved:

Direction	Normalized exact	Mean token F1
English -> Denela	50.0%	90.4%
Turkish -> Denela	43.6%	88.2%
Denela -> English	34.0%	77.8%
Denela -> Turkish	19.2%	61.7%

Exact match is conservative because multiple natural reference translations may preserve the same meaning.

The full training dataset is not included in this model repo. Dataset release/licensing is handled separately from the model weights.

Local Evaluation Snapshot

On a fixed 75-example-per-task validation sample, the promoted v0.2 local models achieved:

Task	Normalized exact	Mean similarity	Mean token F1
English -> Denela	46.7%	87.9%	89.0%
Turkish -> Denela	45.3%	89.6%	88.6%
Denela -> English	50.7%	89.9%	85.8%
Denela -> Turkish	30.7%	81.7%	69.1%

The public app also applies Denela-specific segmentation, punctuation handling, token breakdowns, derivation explanations, and grammar validation around the model outputs.

Evaluation

Evaluation below comes from a held-out 500-example-per-task local evaluation report.

Task	Count	Exact	Normalized Exact	Mean Similarity	Mean Token F1	Invalid Denela
English -> Denela	500	44.4%	61.8%	0.926	0.960	0.0%
Turkish -> Denela	500	44.2%	60.8%	0.917	0.952	0.0%
Denela -> English	500	57.6%	61.0%	0.925	0.939	0.0%
Denela -> Turkish	500	35.6%	41.0%	0.892	0.777	0.0%

The Turkish direction is currently the weakest and may produce fluent but semantically incomplete Turkish, especially for possession, conjunction, and less frequent vocabulary.

Limitations

• Denela is still evolving, so old outputs may become non-canonical.
• The model can confuse semantically close roots when the source sentence is ambiguous.
• Long, nested, or idiomatic sentences may be simplified or partially mistranslated.
• Turkish generation is noticeably weaker than Denela generation and English generation.
• The model learned from synthetic data, so style and coverage reflect the generation pipeline.

Bias, Safety, And Scope

This is a language experiment, not a universal translation model. Because the training data is synthetic and filtered through a small constructed lexicon, it may overrepresent simple declarative sentences and underrepresent real-world cultural, dialectal, or idiomatic language use.

Technical Details

• Base model: google/byt5-small
• Architecture: sequence-to-sequence byte-level T5
• Library: transformers
• Primary checkpoint format: safetensors
• Recommended inference backend: CPU or GPU with torch and transformers
• Default generation used by the Denela API: num_beams=8, max_new_tokens=128

License

The model weights and companion code are released under Apache-2.0.

Training dataset licensing is separate and should be reviewed before publishing the generated dataset itself.