codemaivanngu/dpca-b200-source
dpca-b200-source Tooling that builds, qualifies, packages and publishes the portable B200 runtime for the DPCA reproduction of On-Policy Distillation (ivanniu/On-Policy-Distill). This repo is the harness, not the runtime. The runtime payload (venv, source tree, node helpers) and the model weights live in a separate dataset repo, because they are ~56 GB and the tooling is 61 KB. What is here File What it is dpca-b200-source-20261005.tar.gz The tooling… See the full description on the dataset page: https://huggingface.co/datasets/codemaivanngu/dpca-b200-source.
dpca-b200-source
Tooling that builds, qualifies, packages and publishes the portable B200 runtime for the DPCA reproduction of On-Policy Distillation (ivanniu/On-Policy-Distill).
This repo is the harness, not the runtime. The runtime payload (venv, source tree, node helpers) and the model weights live in a separate dataset repo, because they are ~56 GB and the tooling is 61 KB.
What is here
Verify before using
sha256sum dpca-b200-source-20261005.tar.gz
# 99faca25dd4c5f4660411d3f3ac3278a1fd00561af61e28366b9caddb0cd5ea9
tar -xzf dpca-b200-source-20261005.tar.gz
sha256sum -c dpca-b200-source-20261005.SHA256SUMSProvenance
The overlay is 4 commits on top of the upstream commit. That count is a build-time assertion (OVERLAY_COMMITS_AHEAD = 4), not a claim in this README.
Scope, honestly stated
The tooling proves the runtime executes the paper's objective on the paper's Table 4 model pair, on B200, from a portable payload that another machine can verify byte-for-byte. It does not reproduce the paper's numbers: the paper uses 16 GPUs, 500 steps, response length 16384 and global batch 128. The student SFT checkpoint the paper starts from was never published, and the vLLM rollout path was never qualified.
PROJECT.md in the archive carries the full record, including the defects that only appeared on a second machine.
License
Apache-2.0, matching the upstream repository.
