Reproducible by construction
Content-hashed manifests and clean-tree enforcement make every named run attributable.
MRP
Reproducible LLM training, described by four cards.
MRP — the Model Refinement Pipeline — is a card-driven CLI for LLM training, evaluation, dataset packing, and GPU kernel diagnostics. Four declarative cards plus a content-hashed manifest replace launch-script drift.
Card by card
Four cards in,
one run out.
The pipeline as its console tells it — model, data, train, eval.
$ mrp run --cards model,run,stack,data ✓ manifest sha256:9f2c…e1 clean tree ✓ stack rocm/hip · rdna2 tuned → training step 1200/1200 ppl 8.41 ✓ eval calibration + geometry probes
MRP is a command-line pipeline. Every run is defined by four declarative cards and a content-hashed manifest.
Four cards deep
Runs you can
run again.
If four cards describe a training run completely, any machine can repeat it — that is the whole design.
No brittle launch scripts
Model, run, stack, and data cards own the complete execution contract.
AMD ROCm, first-class
Tuned HIP kernels and stack cards make non-CUDA hardware a deliberate target.
Evaluation in the box
Perplexity, calibration, geometry probes, and diagnostics ship with the training path.
Under the hood
The stack.
The tools and languages MRP is built on.
PythonPyTorchROCm / HIP
More in Technology