LLM Fine Tuning Expert
Posted Updated
Description
Responsibilities:
- Select base models (Qwen, Llama, Gemma, Mistral, Indic models) and adapt tokenizers for Indic and code-mixed text.
- Curate training data from conversations, emails, and financial documents, including synthetic data from larger teacher models.
- Run continued pretraining, SFT, and LoRA/QLoRA or full fine-tuning for text and multimodal models.
- Align models with RLHF, reward modelling, DPO, or GRPO, including verifiable rewards for numeric finance tasks.
- Train at scale on multi-GPU, multi-node setups using FSDP, DeepSpeed, or Megatron with BF16/FP8.
- Design evaluations covering accuracy, numeric hallucination, bias, and robustness; find weaknesses and fix them.
- Build scalable inference pipelines on vLLM, Triton Inference Server, SGLang, or TensorRT-LLM, with quantization for low-latency voice use.
- Work with product, data, and business teams to turn their problems into model specs and success metrics.
Mandatory Skills:
- 3+ years in deep learning with models shipped to production, including at least one model of 1B+ parameters you trained end-to-end on multiple GPUs.
- Hands-on with Hugging Face Transformers, PEFT, and TRL, or equivalents such as Unsloth, Axolotl, or TorchTune.
- Post-training beyond SFT (RLHF, reward modeling, DPO, GRPO, or similar), with measured gains on held-out data.
- Production inference experience on vLLM, Triton Inference Server, or a similar engine.
- Clear written and verbal communication: you can turn training results into a one-page recommendation a business leader can act on.
Preferred Skills:
- Debugging distributed training: divergence, NaN loss, OOM, low GPU utilization, failed resume.
- Building evaluation sets and analyzing model weaknesses such as bias and robustness.
- Hindi, Hinglish, or other Indic language modelling.
- Financial data: bank statements, bureau reports, and credit or collections conversations.
- Knowledge distillation, quantization, and inference optimization.
- Speech or speech-to-speech models.
- Open-source contributions or published checkpoints. • Experience working with business or customer-facing teams.
- Required Skills: Python, AWS Cloud.
Skills
Sign in to read and join the discussion.