tyv-gemma4-e4b-lora

Gemma 4 E4B — Russian ↔ Tuvan translation (LoRA)

LoRA adapter for google/gemma-4-E4B-it fine-tuned for bidirectional Russian ↔ Tuvan translation. Tuvan is a low-resource Turkic language with roughly 280,000 speakers.

On the clean held-out test set this adapter significantly outperforms the specialized translation model Agisight/nllb-rus-tyv-v3-dict_16.5k (NLLB-200-distilled-600M fine-tuned on the same data) in both directions and on every metric.

Results

Clean test set — 1,369 pairs. The held-out test split used for NLLB v3 (shuffle(seed=42), pairs 2000–4000 of Agisight/tyv-rus-200k), with pairs whose Tuvan or Russian side also appears in training removed. Greedy decoding, same length limit as NLLB (32 + 3 × input length new tokens).

Full NLLB test set — 1,999 pairs (comparable with the NLLB v3 model card): chrF++ 49.7 / 49.9 (ru→tyv / tyv→ru) vs. NLLB v3 48.5 / 48.1.

¹ Normalization forgives acceptable variation: for Russian output, words are lemmatized and 3rd-person pronouns merged (Tuvan has no grammatical gender, so «он»/«она» is ambiguous); for Tuvan output, subject pronouns are dropped (person is marked on the verb). ² Unbabel/wmt22-comet-da; reported for tyv→ru only, since COMET's encoder does not cover Tuvan.

Prompt format

The adapter was trained on single-turn chat messages with a short prefix:

• Russian → Tuvan: ru→tyv: <Russian text>
• Tuvan → Russian: tyv→ru: <Tuvan text>

Use exactly this format; other instructions were not seen in training.

Usage

import torch
from transformers import AutoTokenizer, AutoModelForCausalLM
from peft import PeftModel

base_id, adapter_id = "google/gemma-4-E4B-it", "Agisight/tyv-gemma4-e4b-lora"
tok = AutoTokenizer.from_pretrained(adapter_id)
model = AutoModelForCausalLM.from_pretrained(base_id, dtype=torch.bfloat16, device_map="auto")
model = PeftModel.from_pretrained(model, adapter_id).eval()

def translate(text, direction="ru→tyv"):
    prompt = tok.apply_chat_template([{"role": "user", "content": f"{direction}: {text}"}],
                                     add_generation_prompt=True, tokenize=False)
    enc = tok(prompt, return_tensors="pt", add_special_tokens=False).to(model.device)
    out = model.generate(**enc, max_new_tokens=32 + 3 * enc.input_ids.shape[1], do_sample=False)
    return tok.decode(out[0, enc.input_ids.shape[1]:], skip_special_tokens=True).strip()

print(translate("Завтра я поеду в Кызыл к родителям."))
print(translate("Экии!", "tyv→ru"))

Training

• Data: Agisight/tyv-rus-200k, NLLB v3 split; 272,704 training pairs after cleaning, used in both directions — 543,616 examples. Pairs longer than 300 characters were dropped.
• Method: LoRA (r = 32, α = 64, dropout 0.05) on all attention and MLP projections of the language model only — 258 layers, 69.8M trainable parameters (0.87%). Vision and audio towers are untouched. Loss on the translation (completion) tokens only.
• Hyperparameters: 1 epoch (8,494 steps), batch size 64, learning rate 2e-4 with cosine schedule and 3% warmup, max length 256 tokens, bf16.
• Hardware: 1× NVIDIA A100 80 GB (Google Colab), ~7.5 hours.
• Final validation loss: 0.851; validation token accuracy 78.8%. Loss plateaued after ~7,000 steps.

Limitations

• Trained and evaluated on a single test domain; quality on other genres (spoken language, technical texts) is untested.
• Tuvan → Russian sometimes picks a wrong word sense (e.g. «хевис» → «шалаш» instead of «ковёр») and has to guess gender when the Tuvan source does not mark it.
• Tuvan input that uses Н/О/У instead of Ң/Ө/Ү (common when typed without a Tuvan keyboard) can lead to mistranslations.
• The adapter inherits the limitations and license terms of the base model.

Code

Training notebook (Colab), data preparation, and evaluation scripts: Agisight/tyv-translation-mlx.

Credits

• Base model: Google, Gemma 4 E4B.
• Data, fine-tuning, and evaluation: Ali Kuzhuget.