You need to agree to share your Hugging Face information to access this dataset
This repository is publicly accessible, but you have to accept the conditions to access its files and content.
This dataset holds the run artifacts of an ongoing, pre-registered research project (results, logs, saved activations and weights slices). It also contains mirrors of third-party benchmark data under data/, which remain subject to their original licences. Access is granted manually by the author.
Log in or Sign Up to review the conditions and access this dataset content.
sage-decode run artifacts
Artifacts for the sage-decode project: why edits to a code language model's internals rarely change the programs
it writes (late-layer opposition, measured with executable oracles). Models: Qwen2.5-Coder-7B and
DeepSeek-Coder-6.7B-base.
Layout
| path | contents |
|---|---|
f118_run_20260905/ |
every run's outputs, mirroring the project's runs/ directory: per-experiment folders (fXXX/) with results (analysis.json), samples (*.jsonl), logs, and saved tensors (*.npz, *.pt) |
code_20260906/ |
sage-decode-repo.bundle (the git repository, git clone sage-decode-repo.bundle), docs_scripts.tar.gz; pre_squash_20261110/ holds the history before the 2026-11-10 cleanup |
data/ |
mirrors of upstream benchmark data (SecCodePLT, PatchEval, OSV, SALLM, Vul4J, Vul4C) and derived coverage tables; subject to their original licences |
top-level files and experiment_*/ |
early backups from the project's first phase |
Each experiment FXXX is described by docs/PREREG_FXXX.md (pre-registration and deviations) and its entry in
docs/FINDINGS.md, both in the git bundle. Every result is tagged by endpoint; only EXECUTABLE endpoints (verdicts
from running code) carry behavioural claims.
Status
Ongoing research; results are not yet peer reviewed. Please contact the author before citing.
- Downloads last month
- 2