Docs

Open RPT research guide

RPT now publishes the useful path: scripts, notebooks, manifests, results and runtime experiments. The goal is simple: anyone can inspect what worked, reproduce the numbers and build their own variant.

Start here

RPT Repository->Runtime Repository->Model Bundles->README->Reproducibility->Model card->License->Known issues->

Local setup

git clone https://github.com/cesarfavero/RPT.git

cd RPT

python -m venv .venv && source .venv/bin/activate

pip install -r requirements.txt

jupyter lab gemma4_theory/notebooks

Run notebooks in order when reproducing the full trail. For quick inspection, start with the consolidated result files and packed runtime README before launching GPU work.

Linux environment with NVIDIA GPU for the validated runtime path

Python environment from requirements.txt

PyTorch, Transformers and the project runtime dependencies

Enough disk for Gemma-class checkpoints, packed artifacts and result logs

Model license compliance before downloading or redistributing weights

Reproduction path

Run the baseline

Start from the original Gemma 2B reference evaluation and keep the JSON output. This gives the comparison floor used by the later recovery and packed runs.

Open source ->

Prune and recover

Use the progressive pruning, recovery distillation and ternary QAT notebooks to build the student path before expansion.

Open source ->

Expand ternary targets

Compare target layer counts and keep the post-recovery summaries. The working path in the public artifacts converges on the L78 packed recovery line.

Open source ->

Build packed runtime

Generate the packed recovered artifact, validate loaders, then test the native and Triton execution paths before publishing numbers.

Open source ->

Public artifact map

Area	What to use	Link
RPT research	Main public repository with notebooks, scripts, reproducibility docs, result logs and research notes.	Open
Runtime repository	Public runtime surface for builders who want to follow or extend the execution layer separately from the research notebooks.	Open
Model bundles	Bundle-oriented repository for model package organization, manifests, notices and release-side materials.	Open
Core scripts	model.py, bitnet.py and trainer.py hold the compact implementation surface for builders who want to inspect or fork the approach.	Open
Notebooks	The gemma4_theory notebooks document baseline, pruning, recovery, quantization, expansion, export, runtime and benchmark steps.	Open
Runtime	packed_recovered_l78 includes loaders and inference entry points for native and Triton-backed execution experiments.	Open
Results	JSON and CSV outputs preserve the actual benchmark trail, including consolidated summaries, external evaluation and packed recovery logs.	Open
Deployment notes	deploy and ops folders capture package manifests, install notes and release-side operational checks.	Open
Paper and docs	Paper, model card, reproducibility guide, commercialization notes and roadmap keep the research and product rationale auditable.	Open