IndexTeam/Sandglass-Bench

Sandglass: Syllable-Controlled Translation Benchmark

Sandglass is a benchmark for evaluating syllable-controlled translation quality and controllability. It tests whether models can translate Chinese video subtitles into target languages while adhering to precise syllable count constraints.

Overview

• 3,600 test cases: 300 real video subtitles × 4 languages × 3 length targets
• Languages: English (en), Japanese (ja), Spanish (es), Arabic (ar)
• Length targets: short (0.75×), natural (1.0×), long (1.25×) relative to duration-based anchor
• Evaluation: Syllable accuracy + LLM-judged translation quality

Quick Start

pip install -r requirements.txt
hf download IndexTeam/Sandglass-Bench --repo-type dataset --local-dir ./Sandglass-Bench
cd Sandglass-Bench
python prepare_inputs.py --output outputs/model_inputs.jsonl

Send each row's messages to your translation system. Save JSONL predictions as {"case_id": "...", "prediction": "final translation"}; do not send labels or metadata to the model. Preserve the exact prompts and record the model revision and decoding settings.

export JUDGE_API_BASE="https://YOUR_PROVIDER/v1"
export JUDGE_API_KEY="YOUR_API_KEY"
python evaluate.py --predictions outputs/predictions.jsonl --output-dir outputs/evaluation
# Offline syllable-only checks (not full benchmark scores):
python evaluate.py --predictions outputs/predictions.jsonl --output-dir outputs/syllables --skip-judge

The evaluator accepts saved predictions from any inference engine. It does not load a GPU model or bundle provider credentials. The default Judge is gemini-2.5-flash; its request parameters are recorded in the summary.

Data Format

sandglass_bench.jsonl contains 3,600 lines, each with:

{
  "case_id": "uuid:en:natural",
  "prompt": [{"role": "user", "content": "请将下面这句中文视频字幕翻译成英语，译文控制为 12 个音节。\n..."}],
  "label": "reference translation (en only)",
  "metadata": {
    "uuid": "unique-id",
    "fenqu": "category (二次元/影视/旅游出行/游戏/知识)",
    "duration": 1.9,
    "target_language": "en",
    "target_syllables": 12,
    "target_kind": "natural",
    "source": "Chinese subtitle text",
    "summary": "video summary",
    "up_context": "previous subtitle",
    "down_context": "next subtitle"
  }
}

Evaluation Metrics

Syllable Accuracy

• mean|dev|: Mean absolute relative deviation from target
• ±1: Percentage within 1 syllable of target
• dev≤5%/10%/20%: Percentage within relative deviation thresholds
• syllable_reward: Continuous score (1.0 for ±1 or dev≤5%, linearly decaying from 5% to 50%, 0.0 for dev≥50%)

Translation Quality

• quality: LLM judge score (0/0.5/1), evaluates translation correctness without reference
• Judge is informed of duration constraint and tolerates reasonable compression/expansion

Controllability

• slope: Regression slope of (target, output) syllables for same sentence across 3 length targets (ideal: 1.0)
• q_gap_short/long: Quality difference between natural and short/long targets

Final Score

score = (syllable_reward + quality) / 2

Syllable Calculation

The benchmark uses syllable_calculation.py (v3), which supports 22 languages with:

• Reliability varies by language: the source module rates zh/en/ja/ko/vi/es/de/fr/it/pt high, ru/ro/pl/sv/nl/tr/id/ms/ar/fil medium, and th/hi approximate.
• Method: Language-specific rules (pyphen dictionaries, script-based counting, vowel heuristics)
• Number expansion: Digits converted to words before counting

from syllable_calculation import cal_syllable_count

count = cal_syllable_count("Hello world", lang='en')  # 3
count = cal_syllable_count("こんにちは", lang='ja')   # 5

Target Syllable Calculation

Targets are duration-based, independent of Chinese character count:

anchor_syllables = duration_seconds × speech_rate

Speech rates (syllables/second):

• English: 6.19
• Japanese: 7.84
• Spanish: 7.82
• Arabic: 8.4

Three targets per sentence:

• short: round(anchor × 0.75), -1 if collides with natural
• natural: round(anchor × 1.0)
• long: round(anchor × 1.25), +1 if collides with natural

Benchmark Statistics

Dimension	Value
Total items	3,600
Source sentences	300 (5 categories × 60)
Videos	83
Languages	4 (en/ja/es/ar)
Length targets	3 (short/natural/long)
Subtitle duration	0.84-7.0s (median 2.3s)
Chinese chars	10-40 (median 13)
Target syllables (natural)	en: 14, ja: 18, es: 18, ar: 19 (median)

Results Format

results/<model>.jsonl contains per-item results:

{
  "uuid": "...",
  "fenqu": "二次元",
  "target_language": "en",
  "target_kind": "natural",
  "target_syllables": 12,
  "duration": 1.9,
  "source": "如果按照奥特曼世界科技排序的话",
  "hyp": "If we go by the technological hierarchy of the Ultraman world",
  "hyp_syllables": 13,
  "abs_diff": 1,
  "dev": 0.0833,
  "signed_dev": 0.0833,
  "syllable_reward": 1.0,
  "hit_pm1": 1,
  "hit_strict5": 0,
  "hit_10": 1,
  "hit_20": 1,
  "quality": 1.0,
  "judge_error": 0,
  "score": 1.0
}

Evaluation protocol and release changes

All 3,600 rows remain in the denominator. Missing or empty predictions score zero; duplicate and unknown case IDs are rejected. Judge provider failures or unparseable responses abort the run instead of becoming translation scores. --skip-judge and --limit runs are explicitly marked non-formal. Only final translations should be supplied, without reasoning traces or wrapper text.

The published data retains every source prompt, label and scoring field. Only metadata.file_name (video provenance filename) was removed; a stable composite case_id was added. The original syllable-counting modules are unchanged.

This release replaces the source script's GPU generation and positional cache with an engine-independent prediction interface and cache keyed by the complete Judge prompt and request configuration. Strict 0/0.5/1 parsing prevents responses such as 0.1 from silently becoming 1. Subset scores are not full-benchmark results, and scores under a different Judge or generation recipe are not assumed comparable to the historical results below.

Controllability slope is calculated for each source/language over its three length targets, then averaged over groups. Quality gaps are the mean paired short-minus-natural and long-minus-natural quality scores. Per-language and per-length-target metrics, coverage, exact data hash, and Judge parameters are saved to summary.json.

Citation

If you use this benchmark, please cite:

@misc{cui2026homuratamingsandglasstimeconstrained,
      title={HOMURA: Taming the Sand-Glass for Time-Constrained LLM Translation via Reinforcement Learning},
      author={Ziang Cui and Mengran Yu and Chenyu Shi and Yingxuan Shi and Tianjiao Li},
      year={2026},
      eprint={2601.10187},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2601.10187},
}

License

The benchmark data and evaluation code are released under MIT License. The syllable calculation module includes dependencies (pyphen, fugashi, num2words) with their own licenses.

Leaderboard

Top Models

Model	Type	Quality	mean|dev|	±1	dev≤5%	Slope
index-homura	Ours (9B RL)	0.786	0.069	74.4%	50.0%	0.968
gpt-5.6-sol	External API	0.872	0.322	42.5%	28.9%	0.891
Hy-MT2-7B	External OSS	0.856	0.317	18.1%	8.8%	0.195
DeepSeek-v4-flash	External API	0.874	0.337	16.6%	7.5%	0.153
Hy-MT2-30B-A3B	External OSS	0.755	0.274	21.0%	10.6%	0.340
Qwen3.5-9B	External OSS	0.718	0.325	15.6%	6.8%	0.430
Qwen3.5-35B-A3B	External OSS	0.786	0.346	13.3%	5.8%	0.249
Hunyuan-MT-7B	External OSS	0.719	1.058	12.3%	7.4%	0.031
Hy-MT2-1.8B	External OSS	0.590	0.708	14.6%	7.9%	0.222
Qwen3.5-2B	External OSS	0.492	0.515	12.5%	4.5%	0.036

Benchmark collection

Part of the Index-Translate Benchmarks collection.