Skip to content

README.md

Latest commit

 

History

History
56 lines (44 loc) · 2.22 KB

README.md

File metadata and controls

56 lines (44 loc) · 2.22 KB
Raw

IFEval-Audio

Part of the AudioBench benchmark suite. See the main README for installation, the full dataset list, and how to run evaluations.

Overview

IFEval-Audio is a dataset to evaluate instruction-following in audio-based LLMs, with 280 audio-instruction-answer triples across six dimensions: Content, Capitalization, Symbol, List Structure, Length, and Format.

Dataset Structure

  • Audio Input: From Spoken SQUAD, TED-LIUM 3, Muchomusic, etc.
  • Text Instruction: Specifies one dimension (e.g., "Use JSON format").
  • Expected Answer: Reference output.
  • Dimensions: Content, Capitalization, Symbol, List Structure, Length, Format.
  • Distribution: 240 speech triples (40/dimension), 40 music/environmental triples.

Dataset

The dataset is hosted on Hugging Face: YichenG170/AudioLLMInstructionFollowing (you may need to log in and accept the access conditions). It is loaded automatically by the evaluation code — no manual download is required.

How to Run

IFEval-Audio is integrated into AudioBench as the dataset audiollm_instructionfollowing with the metric llama3_70b_judge_combined. From the repo root:

# Step 1 (separate process): serve the Llama-3-70B judge used to score correctness.
bash vllm_model_judge_llama_3_70b.sh

# Step 2: run inference + evaluation for your model.
GPU=0
BATCH_SIZE=1
OVERWRITE=True
NUMBER_OF_SAMPLES=-1               # -1 = all 280 triples

MODEL_NAME=Qwen2-Audio-7B-Instruct # any supported model
DATASET=audiollm_instructionfollowing
METRICS=llama3_70b_judge_combined

bash eval.sh $DATASET $MODEL_NAME $GPU $BATCH_SIZE $OVERWRITE $METRICS $NUMBER_OF_SAMPLES

The judge reports the three metrics below, broken down by the six dimensions.

Evaluation Metrics

IFR: Format adherence score (0/1). SCR: Semantic correctness score (0/1). OSR: Triples with IFR=1 and SCR=1.

Citation

@article{gao2025ifevalaudio,
  title={IFEval-Audio: Benchmarking Instruction-Following Capability in Audio-based Large Language Models},
  author={Gao, Yiming and Wang, Bin and Wei, Chengwei and Sun, Shuo and Aw, AiTi},
  journal={arXiv preprint arXiv:2505.16774},
  year={2025}
}