Large Language Model (LLM) Evaluation Engineer
Umelife · Boon Keng, Central Region ·
- Work mode
- Onsite
- Employment
- Full time
- Category
- AI engineering
Job Responsibilities
- Build and maintain an automated LLM evaluation pipeline covering multiple dimensions, including general capabilities, Agent capabilities, and persona/role-playing. The pipeline should support one-click evaluation, historical result comparison, and regression testing.
- Conduct general capability evaluations using benchmarks such as MMLU, C-Eval, HumanEval, GSM8K, MATH, and IFEval, including benchmark deployment, execution, and results analysis.
- Conduct Agent capability evaluations, including setting up evaluation environments and tracking metrics for benchmarks such as BFCL, τ-bench, and GAIA.
- Design and execute persona/role-playing evaluation frameworks, covering metrics such as identity recognition, role compatibility, multi-turn stability, and style consistency.
- Record and analyze evaluation results from training runs, conduct comparative analysis and anomaly detection, and produce checkpoint evaluation reports.
- Conduct regular intermediate evaluations during the pre-training stage to track the evolution and improvement of model capabilities.
- Bachelor's degree or above in Computer Science, Artificial Intelligence, or a related field.
- Familiarity with mainstream LLM evaluation benchmarks and frameworks, such as lm-eval-harness, OpenCompass, and EvalPlus.
- Strong proficiency in Python, with the ability to independently build evaluation pipelines covering model inference/deployment, batch evaluation, and results analysis.
- Familiarity with LLM inference frameworks such as vLLM and SGLang, with the ability to deploy models for batch inference and evaluation.
- Experience in evaluation data analysis and visualization.
- Detail-oriented and rigorous, with a strong focus on ensuring the reproducibility and reliability of evaluation results.
- Experience with Agent evaluation, particularly BFCL, τ-bench, GAIA, or SWE-bench.
- Experience with persona or role-playing evaluation, such as CharacterBench or RMTBench.
- Experience with evaluation automation and CI/CD integration.
- Understanding of model training workflows, with the ability to understand the relationship between training checkpoints and evaluation results.