qwen3.5-2b-ml-paper-translate

QLoRA fine-tune of Qwen/Qwen3.5-2B 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: 4
• Gradient accumulation: 4
• Effective batch size: 16
• Sequence length: 2048
• Epochs: 1.0
• Learning rate: 0.0001
• QLoRA rank/alpha/dropout: 32/64/0.05
• 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
}

The training dataset, references, and per-example predictions are not distributed from this repository. Only aggregate evaluation results are published.

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": 45.468870997399854,
    "numeric_symbol_preservation": 0.8938547486033519,
    "numeric_symbol_examples": 716,
    "math_preservation": 0.8472222222222222,
    "math_examples": 72,
    "terminology_recall": 0.34812286689419797,
    "terminology_accuracy": 0.34812286689419797,
    "terminology_terms": 293,
    "no_extraneous_prefix": 1.0
  },
  "base-test.metrics": {
    "count": 916,
    "chrf_pp": 44.07591737903462,
    "numeric_symbol_preservation": 0.8801213960546282,
    "numeric_symbol_examples": 659,
    "math_preservation": 0.863013698630137,
    "math_examples": 73,
    "terminology_recall": 0.3767441860465116,
    "terminology_accuracy": 0.3767441860465116,
    "terminology_terms": 215,
    "no_extraneous_prefix": 1.0
  },
  "tuned-valid.metrics": {
    "count": 957,
    "chrf_pp": 66.5621385606034,
    "numeric_symbol_preservation": 0.9175977653631285,
    "numeric_symbol_examples": 716,
    "math_preservation": 1.0,
    "math_examples": 72,
    "terminology_recall": 0.8088737201365188,
    "terminology_accuracy": 0.8088737201365188,
    "terminology_terms": 293,
    "no_extraneous_prefix": 1.0
  },
  "tuned-test.metrics": {
    "count": 916,
    "chrf_pp": 65.9248842697133,
    "numeric_symbol_preservation": 0.9059180576631259,
    "numeric_symbol_examples": 659,
    "math_preservation": 1.0,
    "math_examples": 73,
    "terminology_recall": 0.8372093023255814,
    "terminology_accuracy": 0.8372093023255814,
    "terminology_terms": 215,
    "no_extraneous_prefix": 1.0
  },
  "gpu-stability": {
    "training_completed": true,
    "training_rows": 98127,
    "per_device_train_batch_size": 4,
    "samples": 2633,
    "first_sample": "2026/09/11 15:21:25.431",
    "last_sample": "2026/09/11 19:02:39.979",
    "peak_memory_used_mib": 19785.0,
    "minimum_memory_free_mib": 4308.0,
    "peak_gpu_util_percent": 100.0,
    "peak_temperature_c": 74.0,
    "peak_power_w": 360.9
  }
}

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/qwen3.5-2b-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."},
]
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))

MLX 4-bit on Apple Silicon

hf download Teluteru/qwen3.5-2b-ml-paper-translate --local-dir model-hf
python -m mlx_lm convert --hf-path model-hf --mlx-path model-mlx-4bit   -q --q-bits 4 --q-group-size 64

Use enable_thinking=False; fine-tuning used the non-thinking chat-template prefix. The companion project defaults to the 2B model and can switch between installed 2B and 4B MLX models.

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.