Agent behavior
Curriculum targets decomposition, planning, evidence discipline, tool selection, uncertainty and safe escalation.
PUBLIC · LOCAL AI / ML
Local Reasoning Fabric
A split-host training and deployment workflow for agent behavior, verification, structured work products, memory, SFT and evaluation.
PublicTraining pipelineGemmaOllama runtime
The problem
Why this project exists
Model training and local inference have different infrastructure needs. The repository deliberately separates the training host from the local deployment/runtime host.
Verified capabilities
What is actually implemented
These capability descriptions are intentionally limited to the evidence-backed implementation surface.
SFT
QLoRA SFT, dataset validation and split construction are explicit workflow stages.
RFT
Reward-filtered candidate generation is used instead of uncontrolled self-training.
Evaluation
Immutable evaluation sets cover schema compliance, coding, research grounding, abstention and regression.
Local runtime
Ollama is treated as inference/runtime, not as the trainer.
Safety boundary
Generated data is not treated as the only source of truth and training claims remain bounded.
Architecture
How the system is structured
The diagram below represents the current verified architecture boundary; planned or environment-dependent components are described in the engineering notes rather than silently promoted.
CUDA Training Host
↓
Validated Dataset
↓
SFT / RFT
↓
Immutable Evaluation
↓
Model Artifact
↓
Local Ollama Runtime
Evidence and authority
Seed data and validation/test splits are treated as controlled artifacts.
Base and tuned checkpoints are compared before deployment decisions.
Engineering deep dive
Decisions, trade-offs and boundaries
Technology names matter less than the engineering contracts they enforce.
Seed data and validation/test splits are treated as controlled artifacts.
The documented procedure runs a smaller SFT pilot before larger jobs.
RFT candidates are filtered and reviewed before becoming training data.
Base and tuned checkpoints are compared before deployment decisions.
Project status
Public · evidence-backed status
Training/deployment workflow documented and implemented. Included seed curriculum is explicitly bootstrap-scale, not a production corpus.
ClassificationPublic implementation
EvidenceRepository implementation, tests and project records cross-checked
ClaimsCurrent, planned and limitations kept separate
DeploymentNot changed by portfolio integration
PUBLIC IMPLEMENTATION
Inspect the implementation directly.
Public implementation — source and training workflow are available.