Qwen3.5 Technical English-to-Japanese Translator

QLoRA fine-tune of Qwen/Qwen3.5-4B-Base for deep learning, robotics, language models, and vision-language-action research text. Training used assistant-only loss and a locally generated, research-grounded synthetic corpus.

This directory contains the LoRA weights merged into the full Hugging Face base model. It is the production model, not a smoke-test artifact.

Training

• Training rows: 98127
• Per-device batch size: 2
• Gradient accumulation: 8 (effective batch size 16 on one GPU)
• Sequence length: 2048
• Epochs: 1
• Learning rate: 1e-4
• QLoRA: NF4 4-bit, rank 32, alpha 64, dropout 0.05
• Training hardware: NVIDIA GeForce RTX 4090 24 GB
• Objective: English-to-Japanese causal language modeling with assistant-only loss
• Target domains: deep learning, robotics, language models, and vision-language-action models
• Grounding: recent arXiv metadata, abstracts, and available ar5iv excerpts

Dataset

{
  "accepted": 100000,
  "rejected": 24871,
  "splits": {
    "train": 98127,
    "valid": 957,
    "test": 916
  },
  "domains": {
    "deep_learning": 25123,
    "language_models": 24548,
    "robotics": 25686,
    "vision_language_action": 24643
  },
  "complete": true
}

Evaluation

chrF++ is reported on a 0–100 scale. Preservation and terminology values are fractions from 0–1.

{
  "base-valid.metrics": {
    "count": 957,
    "chrf_pp": 38.796437754138005,
    "numeric_symbol_preservation": 0.8868715083798883,
    "numeric_symbol_examples": 716,
    "math_preservation": 0.9861111111111112,
    "math_examples": 72,
    "terminology_recall": 0.4334470989761092,
    "terminology_accuracy": 0.4334470989761092,
    "terminology_terms": 293,
    "no_extraneous_prefix": 1.0
  },
  "base-test.metrics": {
    "count": 916,
    "chrf_pp": 37.8100085836244,
    "numeric_symbol_preservation": 0.8861911987860395,
    "numeric_symbol_examples": 659,
    "math_preservation": 0.9315068493150684,
    "math_examples": 73,
    "terminology_recall": 0.4,
    "terminology_accuracy": 0.4,
    "terminology_terms": 215,
    "no_extraneous_prefix": 1.0
  },
  "tuned-valid.metrics": {
    "count": 957,
    "chrf_pp": 66.65652158010008,
    "numeric_symbol_preservation": 0.9245810055865922,
    "numeric_symbol_examples": 716,
    "math_preservation": 1.0,
    "math_examples": 72,
    "terminology_recall": 0.8020477815699659,
    "terminology_accuracy": 0.8020477815699659,
    "terminology_terms": 293,
    "no_extraneous_prefix": 1.0
  },
  "tuned-test.metrics": {
    "count": 916,
    "chrf_pp": 66.47080563825857,
    "numeric_symbol_preservation": 0.9089529590288316,
    "numeric_symbol_examples": 659,
    "math_preservation": 1.0,
    "math_examples": 73,
    "terminology_recall": 0.8465116279069768,
    "terminology_accuracy": 0.8465116279069768,
    "terminology_terms": 215,
    "no_extraneous_prefix": 1.0
  },
  "gpu-stability": {
    "training_completed": true,
    "training_rows": 98127,
    "per_device_train_batch_size": 2,
    "samples": 6080,
    "first_sample": "2026/09/07 17:02:20.626",
    "last_sample": "2026/09/08 01:33:11.387",
    "peak_memory_used_mib": 24090.0,
    "minimum_memory_free_mib": 3.0,
    "peak_gpu_util_percent": 100.0,
    "peak_temperature_c": 75.0,
    "peak_power_w": 365.58
  }
}

Local deployment

Convert this merged model to MLX 4-bit on Apple Silicon with qwen-translate export-mlx, then install the Native Messaging Host with scripts/install_macos.sh. The companion Manifest V3 Chrome extension translates selected text locally and restores ar5iv MathML after generation.

Transformers

import torch
from transformers import AutoModelForImageTextToText, AutoTokenizer

model_id = "Teluteru/qwen3.5-4b-ml-paper-translate"
tokenizer = AutoTokenizer.from_pretrained(model_id)
model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype=torch.bfloat16,
    device_map="auto",
)
messages = [
    {
        "role": "system",
        "content": (
            "Translate technical English into natural Japanese. Preserve equations, "
            "numbers, units, identifiers, and formatting. Return only the translation."
        ),
    },
    {
        "role": "user",
        "content": "The VLA policy predicts robot actions from images and instructions.",
    },
]
prompt = tokenizer.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
    enable_thinking=False,
)
inputs = tokenizer(prompt, return_tensors="pt").to(model.device)
with torch.inference_mode():
    generated = model.generate(**inputs, max_new_tokens=512, do_sample=False)
print(tokenizer.decode(generated[0, inputs.input_ids.shape[1]:], skip_special_tokens=True))

Use enable_thinking=False; the model was trained with the non-thinking chat-template prefix.

MLX 4-bit on Apple Silicon

python -m pip install -U mlx-lm huggingface-hub
huggingface-cli download Teluteru/qwen3.5-4b-ml-paper-translate \
  --local-dir qwen3.5-4b-ml-paper-translate-hf
python -m mlx_lm.convert \
  --hf-path qwen3.5-4b-ml-paper-translate-hf \
  --mlx-path qwen3.5-4b-ml-paper-translate-mlx-4bit \
  -q --q-bits 4 --q-group-size 64

The repository publishes merged BF16 Hugging Face weights. MLX conversion must be run on an Apple Silicon Mac; measured M3 load time, memory, and throughput will be added after the target machine acceptance run.

Evaluation protocol

• Validation: 957 examples; Test: 916 examples.
• chrF++ uses SacreBLEU CHRF(word_order=2).
• Numeric/symbol preservation is a strict sentence-level exact-multiset match for protected numeric tokens and identifiers.
• Math preservation is a strict sentence-level exact-multiset match for inline/display math and [[MATH_NNNN]] placeholders.
• Terminology accuracy is exact occurrence of the expected Japanese glossary term.
• Base and tuned models use deterministic greedy decoding with thinking disabled.

Aggregate machine-readable metrics and Base/Tuned comparisons are uploaded under evaluation/. Evaluation source texts, references, and generated predictions are intentionally not published.

Limitations

• The corpus and references are synthetic and can contain unnatural wording or factual errors.
• Automatic preservation scores are not perfect; callers should retain deterministic MathML placeholder checks and verify numbers for consequential use.
• Qualitative review found that unsupervised/supervised may occasionally be rendered as the understandable but unnatural 非監督/監督; 教師なし/教師あり is preferred.
• This model is not validated for safety-critical robot control or languages other than English to Japanese.

The data and automatic metrics may contain errors. Review translations before safety-critical use. MathML on ar5iv pages is preserved by the companion Chrome extension outside the language model.

Citation

@misc{teluteru2026qwen35technicalja,
  title        = {Qwen3.5-4B ML Paper Translate},
  author       = {Teluteru},
  year         = {2026},
  howpublished = {Hugging Face model repository},
  url          = {https://huggingface.co/Teluteru/qwen3.5-4b-ml-paper-translate}
}