AI 大模型算法工程师（后训练 / 领域适配方向）
北京、上海、深圳
社招
全职
技术
职位 ID：A02467
职位描述
你将参与 MiniCPM 系列及相关大模型在政企真实场景中的能力适配，让模型在对应领域的长文档、知识问答、复杂推理、多模态理解、智能体执行和端侧部署等任务上持续变强。这里的算法工作与客户业务紧密相连，模型效果最终要经得起真实数据、真实流程和真实设备的检验。你可能参与领域继续训练、SFT、LoRA/QLoRA、蒸馏、偏好优化、强化学习、数据合成、难例挖掘、量化和推理优化，也可能深入某个行业场景，把模型能力与知识库、工具和业务规则组合起来。我们期待你有研究视野，也愿意把实验结果落到产品和项目里。 岗位职责 1. 负责大语言模型、多模态模型或智能体模型的训练、后训练、领域适配和效果优化。 2. 围绕政务办公、知识库问答、情报分析、工业质检、数字员工、长文档理解和端侧多模态等场景，定义算法问题并设计实验方案。 3. 参与数据采集、清洗、标注、合成、质量筛选、难例挖掘和训练数据配比，推动数据、算法、评测和应用形成闭环。 4. 使用 PyTorch、Transformers、LLaMA-Factory、verl、DDP/FSDP 或其他训练框架完成训练和调试，分析 loss、能力变化和失败案例。 5. 参与提示词、检索、工具调用、规划、反思、奖励建模和 Verifier 等模块的算法优化，提升模型在复杂任务中的稳定性和泛化能力。 6. 参与模型量化、蒸馏、剪枝、投机采样、算子适配和推理加速，建立可复现的实验记录、模型版本、数据版本和评测结果，将有效方法转化为团队能力。
职位要求
投递
ModelBest established in August 2022, is an Artificial Intelligence technology company headquartered in Beijing, China. The company is deeply rooted in the field of general AI, with a focus on the innovation and application transformation of large-scale models. ModelBest boasts a prestigious founding R&D team from the Tsinghua in the AI field. Leveraging numerous cutting-edge technologies in natural language processing, the company is currently constructing a large-scale pre-trained model library and corresponding tools, aiming to standardize the technology and applications of large models.
Based on the capabilities of the CPM series of large models, ModelBest has successfully facilitated the intelligent upgrade and efficiency enhancement of various industries. Moreover, the company has released a multi-modal dialog assistant with hundreds of billions of parameters, “Luca", to the public. Adhering to the core philosophy of "intelligence encompassing all", we look forward to exploring and expanding unknown possibilities with global partners, aiming to ensure AI technology serves humanity for a better life in a safe and inclusive manner, laying a solid foundation for the advent of the AGI world.
OpenBMB (Open Lab for Big Model Base，https://github.com/OpenBMB), founded by ModelBest Inc & TsinghuaNLP, aims to build foundation models and systems towards AGI.
