Initialize a workspace
tiny-prl.json and writes a working algo.py. It does not overwrite an existing algorithm file.
The SmolLM2 choices prepare their applicable packaged decode routes. Qwen uses stock vLLM. Selecting an environment does not establish that the selected model can learn it; each workload still needs a baseline and a resource check.
Train with the edited algorithm
--trainer.optim.lr 2e-5. Training from an initialized workspace imports algo.compute_advantages. The function file is also copied into the run’s configs for provenance.
Upload is the cleanup gate
The default naming rule isPradheep1647/tiny-prl-<model>-<env>-rNN on this account, with the next free run number chosen at training start. For example, tiny-prl-qwen3-0.6b-gsm8k-r01. An explicit --hf-repo owner/name overrides it. New repositories default to private.
After successful training, the uploader exports the final adapter or full model. It verifies each remote file’s content identity at the immutable upload commit. Only then does it remove the run-owned base model, broadcasts, local checkpoints and export. It retains small logs, metrics, configs and the HF link. The HF export omits optimizer state for exact resume.
If upload or verification fails, weights stay local. Retry with:
What must remain installed
~/.local/bin/tiny-prl calls the roughly 31 MiB runtime under ~/.local/share/tiny-prl. It reuses ~/.venv, its uv-managed Python, CUDA/nvcc and NVIDIA drivers. It is not a bundled replacement for that stack.
Base weights come from HF and can be downloaded again. Kernel source lives in the installed runtime; cleared build caches can be compiled again. The original large research checkout can be archived separately.
Launcher · Uploader and deletion gate · Verified upload/cleanup records
Continue with the code map.