The Dataset Viewer has been disabled on this dataset.
bilbo retrieval eval
The data of bilbo's retrieval eval: which memory strategy best remembers what happened in real coding sessions? The code, and how each file here was made, is in evals/retrieval.
corpus-<sha12>.tar.zst: 403 real coding sessions from twenty-two public projects, untouched, one file per session.<sha12>starts the archive's sha256. The first five projects were chosen by metadata, the other seventeen by a screen for procedure.corpus-1235aec0592a.tar.zstholds the first five alone (145 sessions).store-<arm>-<key>.tar.zst: the memory one strategy built from the corpus.<key>hashes what shapes it: the corpus, the embedder, the LLM for an arm that uses one, and the arm's files.
The eval's scripts download each file at the commit recorded in the eval's dataset.lock and check its sha256 against
the lock, so a run always gets the bytes it was pinned to.
Credits
The sessions come from SWE-chat (Baumann et al., SWE-chat: Real-World AI Coding Sessions in the Wild, COLM 2026), released under ODC-BY 1.0.
- Downloads last month
- 185