Yadia / Compute
Compute partnerships
Yadia is an independent research lab. Compute decides how many questions we can answer, and how firmly. We welcome research credits and compute grants from cloud and GPU providers.
What the compute is for
We post-train open multimodal models with reinforcement learning to reason about cause and effect over images and text, starting with medical imaging such as CT. Alongside that, we test by intervention whether a model’s stated reasoning causes its answer, and measure what models represent internally.
Every reinforcement-learning run is paired with a matched control that receives the same GPU time without reinforcement learning. Hypotheses, pass criteria and held-out test sets are fixed before training starts. A claim therefore costs at least two full training runs plus evaluation.
How we run today
- On-demand single-GPU cloud instances and local workstation GPUs.
- Open-weight models, fitted to the hardware with parameter-efficient fine-tuning (LoRA) and quantised weights.
- Reinforcement-learning workloads: many sampled rollouts per problem, followed by gradient passes.
What more compute unlocks
- Larger models. Testing whether results hold beyond the small open models we can fit today.
- Full-precision training. Removing quantisation as a possible confound.
- Repeat runs. Several seeds per configuration, so results come with variance, not single runs.
- More controls and ablations. Ruling out alternative explanations one change at a time.
- Faster turnaround. Experiments that now take days of sequential runs, in parallel.
Useful hardware
- High-memory data-centre GPUs (80 GB class and above) for single-node reinforcement learning.
- Multi-GPU nodes for larger models and parallel arms of the same experiment.
- Modest persistent storage for checkpoints and evaluation outputs.
What partners get
- Acknowledgement in every write-up the compute supports.
- Public write-ups of methods and results on this site, including results that do not confirm our hypotheses.
- Practical feedback on the platform from a reinforcement-learning workload.
- A case study, by agreement.
Contact
For credits, grants and research partnerships, email hello@yadia.tech.
Yadia is operated by Myelin Limited.