MichiganNLP/language-energy-divide

🌍⚡ The Language–Energy Divide

Per-language energy measurements & prompts for multilingual LLM inference

This dataset accompanies the paper "The Language–Energy Divide: Measuring Energy Costs of Multilingual LLM Inference." It releases the per-language energy measurements and the prompts used in the study, so researchers can build on our numbers without re-running the full sweep (≈ the entire measurement campaign).

⚡ Per-token energy varies by up to 8.3× across languages · 🔋 total energy per request set varies by up to 179× (English 17.6 kJ → Pashto 3,147 kJ) · 📉 the most energy-expensive languages are also the least accurate.

All energy is measured with the ML.ENERGY Benchmark (vLLM serving + the Zeus library) and reported as steady-state per-token energy. Hardware: NVIDIA L40S (main / cross-model / cross-task) and RTX 6000 Pro Blackwell (batch-size sweep).

📦 Contents

📊 results/ — energy measurements

File	Description
belebele_canonical_Qwen3-8B_0shot.csv	⭐ Main result. Qwen3-8B, Belebele, 122 languages: energy/token, output tokens, total energy, accuracy (strict & lenient), resource level.
belebele_canonical_Qwen3-14B.csv, …Qwen3-8B.csv, …Llama-3.1-8B.csv, …Llama-3.1-70B.csv	Per-language energy for additional models (cross-model study).
cross_model_12lang_table.csv	Cross-model per-token energy comparison.
crosstask_seqs256_summary.csv	Cross-task (Belebele / GSM8K / LM-Arena) per-token energy, 8-language subset.
belebele_seqsweep_l40s_0shot.csv	Batch-size sweep on L40S (max_num_seqs ∈ {16,32,64,128,256,512}), 8 languages.
seqsweep_belebele_v1_qwen8b_RTX6000_all8.csv	Batch-size sweep on RTX 6000 Pro Blackwell, 8 languages.

Key columns: steady_energy_per_token_J (J/token), total_output_tokens, whole_gpu_energy_J, accuracy_strict / accuracy_lenient, resource_level (NLLB high/low).

📝 prompts/ — the data used

Path	Description
belebele_instructions_all_languages.json	Per-language zero-shot CoT instructions + primers for all 122 Belebele languages (NLLB-200 translated, back-translation QC, hand-curated where needed).
belebele_instruction_manual.csv	Human-readable table of the per-language instruction/primer translations + curation notes.
gsm8k/gsm8k_<lang>_scored.jsonl	Machine-translated GSM8K math prompts with back-translation and quality scores (bertscore_f1, comet).
lmarena/<lang>_scored.jsonl	Machine-translated LM-Arena open-chat prompts with back-translation and quality scores.

Each translated prompt record: id, prompt_en (source), prompt_tgt (translation), back_en (back-translation), bertscore_f1, comet. Translations passing the quality bar (BERTScore ≥ 0.85, COMET ≥ 0.75) were retained.

🌐 The 8-language subset

Four high-resource — English (Latin), Chinese (Hans), Russian (Cyrillic), French (Latin) — and four low-resource — Southern Pashto (Arabic), Tigrinya (Ethiopic), Shan (Myanmar), Tibetan (Tibetan) — used for the cross-model, cross-GPU, and cross-task experiments.

🧪 Quick start

import pandas as pd
df = pd.read_csv("hf://datasets/MichiganNLP/language-energy-divide/results/belebele_canonical_Qwen3-8B_0shot.csv")
df.sort_values("steady_energy_per_token_J").head()

📚 Citation & attribution

@article{language-energy-divide,
  title  = {The Language--Energy Divide: Measuring Energy Costs of Multilingual LLM Inference},
  author = {Deng, Naihao and Shen, Alissa and Feng, Yiming and Nwatu, Joan and
            Chung, Jae-Won and Chowdhury, Mosharaf and Chen, Yulong and Mihalcea, Rada},
  year   = {2026}
}

Built on: Belebele (Bandarkar et al., 2024, CC-BY-SA-4.0), GSM8K (Cobbe et al., 2021, MIT), LM-Arena / Chatbot Arena (Chiang et al., 2024), translation via NLLB-200 (NLLB Team, 2022), energy measurement via the ML.ENERGY Benchmark (Chung et al., 2025). Translated prompts derive from these sources; please also cite them and respect their licenses.

⚠️ Note on translations

GSM8K and LM-Arena prompts, and the Belebele per-language instructions/primers, are machine-translated (with QC), not natively authored. They may not fully reflect real-world usage of each language. Measured disparities are best read as a lower bound on what speakers of underserved languages face with natively authored inputs.