ML Researcher, Foundational Models

This is the research seat that decides what Sarvam's next foundation models look like, and the posting is unusually direct about what it is not: not a fine tuning role, not an apply known recipes role. You take a vague hunch about architecture, optimisation, data composition or training dynamics, design the right ablations, run them at meaningful scale, and turn the result into a decision that ships in a production training run. You get direct access to large compute and a tight loop with the engineers building the training stack and data pipelines, and you are expected to disagree with the team and defend your reasoning with evidence. The bar is high and specific: a PhD, first author publications at NeurIPS, ICML, ICLR, ACL, EMNLP or COLM, and hands on experience pre training transformer models from scratch, ideally at 7B parameters or more.

What the posting requires:
- A PhD in Machine Learning, Computer Science or a closely related field, or in the final stages of completion.
- 3+ years of research experience post PhD, or equivalent depth. The posting adds that exceptional early career candidates with a strong research record will be considered, so the years figure is not an absolute gate.
- First author publications at top tier ML venues such as NeurIPS, ICML, ICLR, ACL, EMNLP or COLM.
- Hands on experience pre training transformer based language models from scratch, ideally at 7B parameters or more. The posting says you should be able to describe a training run you owned end to end, including what went wrong and how you debugged it.
- Meaningful contributions to the open source LLM ecosystem: research code, model releases, datasets or substantive contributions to widely used projects.
- Fluency in PyTorch and comfort with distributed training. You should be able to read a training loop and immediately see where it might be slow, unstable or wrong.
- Strong intuition for experimental design: knowing what to measure, what to ablate, and what scale a result needs to hold at before it can be trusted.

Nice to have:
- Work on novel architectures such as mixture of experts, state space models or hybrid architectures, or non trivial modifications to standard transformers.
- Experience with multilingual or multimodal pre training.
- Research contributions in post training: RLHF, RLVR, distillation or reasoning models.
- A track record of taking a research idea from prototype to a shipped capability.

What the work actually looks like:
- Drive open ended research on architecture, optimisation, scaling behaviour, training stability and post training recipes for the next generation of foundational models.
- Design and execute ablations at scales that genuinely inform large run decisions, including running pre training experiments end to end yourself.
- Translate findings into concrete proposals for the next training run, and own those proposals through to production.
- Work directly with the infrastructure and data teams, since many of the important research questions live at that boundary.
- Read broadly, write internally, and publish externally when the work merits it.

Why the role is framed as rare: the posting says you would be one of a small number of people in the world making architectural and training recipe decisions on a frontier scale model run, with the compute and autonomy to back it up.

Location and working pattern: Bengaluru. No remote or hybrid arrangement is stated.

Honest fit guidance: the 3 year figure is genuinely low for the seniority of the work because it is counted after a PhD. If you do not have a PhD and a first author publication record, this specific posting is not the one to spend an application on, whatever your industry experience.