Hugging Face's logo

Datasets:
EfficientLLMInferenceCompetition
/
efficiency_samples_10pct

Tasks:
Text Generation
Languages:
Arabic
Persian
Urdu
Size:
n<1K
License:
Dataset card Data Studio Files
xet
 Community
efficiency_samples_10pct / README.md
omidgh's picture
omidgh
Add public 10% stratified efficiency-eval sample
49a6c76 verified
preview code
|
Raw
History Blame Contribute Delete
3.71 kB
metadata 
pretty_name: Muslim Languages Efficiency Eval Samples (10%)
license: other
license_name: mixed-apache-2.0-cc-by-sa-4.0
language:
  - ar
  - fa
  - ur
  - bn
  - id
task_categories:
  - text-generation
size_categories:
  - n<1K

Muslim-language LLM efficiency-eval samples (public 10%)

Public 10% stratified sample of the private dataset EfficientLLMInferenceCompetition/efficiency_samples.

Workload for the NeurIPS 2026 competition proposal Efficient LLM Inference for Diverse Muslim Languages and Cultures.

This repo holds constructed GuideLLM Poisson traffic (not model weights). Lengths are measured with the official google/gemma-4-31B-it tokenizer including the chat template.

Sampling: seed 42, 200 conversations (40 per language), preserving the official mix (20/25/15/15/10/15). Some mixed conversations reuse a unique cell, so processed/ has 184 cells.

Files

Path Description
mixed/conversations.jsonl Primary set: 200 conversations (GuideLLM conversation_turns)
mixed/sidecar.jsonl example_id → language / workload join keys
mixed/manifest.json Mix counts, sample seed, and notes
mixed/single_turn.jsonl 170-row convenience split
mixed/multi_turn.jsonl 30-row convenience split
processed/{ar,fa,ur,bn,id}/*.jsonl Unique cells used by this sample (184)
configs/benchmark.yaml Frozen protocol (240 s windows, SLOs, λ fractions)
NOTICE Third-party attribution (Aya Apache-2.0, Wikipedia CC BY-SA 4.0)

Languages (equal 20%)

Code Language Register n
ar Arabic MSA only (dialects out of scope) 40
fa Persian Formal; preserve ZWNJ 40
ur Urdu Native script only (Roman Urdu out of scope) 40
bn Bengali Standard Bengali 40
id Indonesian Bahasa Indonesia only 40

Workload mix (experimental, not production traffic)

Workload Input tokens Output budget Mix n
short_interaction 64–512 16–128 20% 40
ordinary_single_turn 256–2048 128–512 25% 50
long_generation 256–2048 1024–4096 15% 30
long_context_short 8192–32768 64–512 15% 30
long_prompt_long_generation 8192–32768 1024–4096 10% 20
multi_turn_replay 3–8 turns; 512–16384 accumulated 64–512 / turn 15% 30

output_tokens_count is a credit cap (max_tokens), not a forced length.

Sources

Source URL License Role
Aya (train) https://huggingface.co/datasets/CohereForAI/aya_dataset Apache-2.0 Short / ordinary seeds
Wikipedia (ar, fa, ur, bn, id) MediaWiki API CC BY-SA 4.0 Long documents and long-context bundles

Redistribute Wikipedia-derived text under CC BY-SA 4.0 with attribution. See NOTICE. Each mixed row has a license field (Apache-2.0 or CC BY-SA 4.0).

Protocol (summary)

  • Measurement window: 240 s (plus 15 s warmup / 15 s cooldown)
  • SLOs: TTFT < 10 s, decode ≥ 20 tok/s
  • SUT = successful-user output tokens / window seconds (failed UX → 0)
  • Score = mean(SUT_L, SUT_M, SUT_H) at 0.3 / 0.6 / 0.9 × λ_ref
  • Arrivals: Poisson via GuideLLM

Load 

from datasets import load_dataset

ds = load_dataset(
    "EfficientLLMInferenceCompetition/efficiency_samples_10pct",
    data_files="mixed/conversations.jsonl",
    split="train",
)