From the source
Summary by forck
Together AI introduced Together Post-Training, a service for building custom models on its platform through reinforcement learning from environment rewards and supervised fine-tuning on supervision data.
- The company says the service lets customers define the training loop while Together co-optimizes the systems and infrastructure underneath for model quality and training speed.
- Together argues that training custom models is no longer niche, as companies seek to cut costs and improve task-specific performance, and that many that could benefit never start or stall because post-training demands repeated work on data curation, training algorithms and evaluation.
- Experiments run on Model Resources, dedicated GPU allocations that are not shared with other users' requests, and each resource can hold one or several Training Sessions tied to a single base model.
- Together says users can run LoRA training, including multiple LoRA adapters sharing the same GPU capacity, as well as full-weight experiments for select models, and can start a model resource with zero generators for SFT to save costs.
