Pre-trained data researchers/R&D engineers

Pre-trained data researchers/R&D engineers

Share

Full-time | Full-stack development and algorithms | DeepSeek|Zhejiang · Hangzhou City, Beijing City

Apply for a position

Job description

As a pre-trained data researcher/R&D engineer, you will be a direct driver of core model iterations. In daily work, we reshape large model cognition through data, exploring the limits of Data-centric AI's capabilities.

[Positions]

1. Data Quality and Governance

• Design and execute large-scale training data cleaning processes, build a data quality evaluation system, and continuously optimize data allocation and screening strategies

• It studies the optimal learning mechanisms of pre-training data quality, diversity, and repetition with model scaling, exploring maximizing data benefits

• Collaborate deeply with upstream and downstream partners to formulate labeling standards and quality standards, and build a closed-loop feedback mechanism

2. Model ablation and evaluation

• Ablation experiments were designed for dimensions such as data formulation, data quality, and data scale, quantifying the impact of data changes on model capability

• Build a multidimensional model evaluation system (automatic + manual), covering general capabilities, vertical scenarios, and security

3. Vertical capability research

• Multimodal Data: Participate in the collection, cleaning, and quality control of multimodal training data such as images and text, audio, and video

• Search and vector recall: Optimize the quality of retrieval data in RAG scenarios, participate in embedding model training data construction and recall evaluation effectiveness

• Emotional intelligence and intention understanding: Studying data strategies related to emotional intelligence and intention understanding

4. Cross-team collaboration and co-construction of standards

• Collaborate with model strategy product managers to build data-model-product quality standards, driving data-driven model iteration and closed loops

• Transform business requirements into data requirements, coordinating resources for R&D, annotation, and product distribution

[Job Requirements]

1. Basic requirements

• CS/AI-related major, bachelor's degree or above, research or work experience related to pretrain/posttrain data and other large-scale data such as information retrieval

• Familiar with mainstream large model training processes (Pretrain / SFT / RL / OPD) and understand the driving mechanisms of data on model behavior

• Proficient in Python and any common data processing toolchains (pandas / Spark / Duckdb, etc.)

• Possesses solid experimental design and statistical analysis capabilities, able to independently design rigorous ablation experiments and draw reliable conclusions

• Background in pre-training data, information retrieval direction, and multimodal research

(Requires at least 2 items)

2. Extra points

• Curious about the world and having a unique understanding of model behavior,

• Possess a good research taste, be good at thinking and identifying the low points in jagged intelligence

• Excellent cross-team communication skills, enabling efficient collaboration with multiple roles such as annotation, product, and engineering

• Publications at top conferences or contributions to core models

• Experience in implementing large-scale data cleaning pipelines (TB-level tokens)

Job information

Job title

Pre-trained data researchers/R&D engineers

Salary range

-

Nature of the position

Full-time

Types of functions

Full-stack development and algorithms

Department

DeepSeek

Work location

Zhejiang · Hangzhou City, Beijing City