Чётко очерченная исследовательская роль в заметной AI-лаборатории даёт возможность влиять на решения, которые внедряются в крупные модели, но высокий порог требований и регулярные поездки во Фрайбург делают вакансию довольно специализированной.
?Очень высокий порог
Передовой ИИ
Влияние на продукт
Современный стек
?Нужно бывать в офисе
Выше рынка
Мэтч & Сопровод
Для мэтча с этой вакансией нужен Plus
Подключить Plus
Описание вакансии
Текст:
Hirify AI/Оригинал
TL;DR
Member of Technical Staff - Safety (AI): Developing and shipping safety methods for open-weight FLUX image and video generation models with an accent on safety training, large-scale data filtering, concept erasure, and multimodal evaluation. Focus on designing robust defenses against harmful generations, adversarial prompts, and downstream fine-tuning while preserving image quality and prompt following.
Location: Freiburg, Germany; attendance at the Freiburg office at least 2 days a week or one full week every other week
Compensation: EUR 160K–220K plus equity, depending on profile
Company
Black Forest Labs develops generative image and video models, including Latent Diffusion, Stable Diffusion, and FLUX, with a focus on research excellence, open science, and human creativity.
What you will do
Develop mid- and post-training safety methods for FLUX, including safety SFT, preference optimization, reinforcement learning with safety rewards, and concept erasure.
Build classifiers and data-filtering pipelines operating at pretraining scale across billions of images, video clips, and captions.
Improve safety for open-weight releases, including robustness against adversarial prompts and downstream fine-tuning.
Extend safety methods to prompts, captions, and text encoders conditioning image and video models.
Design evaluations and red-teaming to measure harmful generations alongside image quality and prompt following.
Work with pretraining and post-training teams to ship safety methods in release models and publish research where appropriate.
Requirements
Hands-on research experience in visual safety for text-to-image, text-to-video, or vision-language models, including harm classifiers, safe generation methods, or safety benchmarks.
Experience with either large-scale safety data filtering over hundreds of millions of samples or more, or safety mid-/post-training using safety SFT, DPO, reinforcement learning objectives, or unlearning for large generative models.
Deep familiarity with concept erasure methods such as ESD, UCE, and adversarial unlearning, including concept recovery and collateral damage failure modes.
Track record of shipping through top-venue publications paired with production impact or clear production wins at a frontier lab.
Experience with distributed training at multi-node scale.
Strong PyTorch fluency and the ability to write research code that others can build on, verify coding-agent output, and own the result.
Experience running multi-agent workflows and deciding which tasks to delegate or retain.
Nice to have
Experience maintaining safety after open-weight release and downstream fine-tuning.
Experience with video-generation safety, audio safety for jointly generated audio and video, text safety, multilingual prompt and jailbreak defenses, or text safety classifiers.
Experience unlearning concepts from text encoders or building safety-tuned encoders.
Experience with latent-space safety methods such as activation steering or safe guidance.
Experience building VLM-based safety judges or guard models for filtering or evaluation.
Culture & Benefits
Work in a frontier research lab focused on deep understanding, ambitious research, and shipping production technology.
In-person collaboration in Freiburg is required at least 2 days per week or one full week every other week.
Reasonable travel costs for the required Freiburg attendance are covered.
Values include research rigor, low ego, bold execution, shared credit, and kindness.
Compensation includes EUR 160K–220K plus equity, depending on profile.
Будьте осторожны
Если работодатель просит войти в их систему, используя iCloud/Google, прислать код/пароль, запустить код/ПО, не делайте этого. Это мошенники. Обязательно жмите «Пожаловаться» или пишите в поддержку. Подробнее в гайде →