A DeepSeek-R1 variant was fine tuned for smart contract vulnerability detection using GRPO and LoRA on a single A100 80GB. The result is a task specific model optimized for screening, not full audits.
Key outcomes:
strong improvement in vulnerable vs clean classification
reliable structured outputs, better integration into pipelines
weak performance in deep vulnerability reasoning, SWC identification remains limited
The model learned formatting faster than security understanding. Usability improved before trustworthiness.
Role of Decentralized Compute
@Fluence enabled the entire training run at a cost of $30.97.
This is the shift:
GPU access without centralized cloud pricing
on demand compute sourced from distributed providers
low cost experimentation with specialized models
Outcome
Instead of expensive general models, developers can iterate on cheap, targeted systems for specific tasks like first pass contract screening.
Fluence reduces compute cost.
Lower compute cost enables rapid, niche AI development.