The dataset viewer is not available because its heuristics could not detect any supported data files. You can try uploading some data files, or configuring the data files location manually.
System One training pairs
Everything needed to fine-tune the System One cross-encoder on a GPU box, with no git clone.
| path | what it is |
|---|---|
data/pairs-gold.jsonl |
compiled (premise, hypothesis, label) pairs from each dataset's own labels |
data/distill/*.jsonl |
Claude Haiku 4.5's answers next to gold, one row per training item |
data/cache/*.jsonl |
the deterministic train/calib/test splits, seed 42 |
src/ |
systemone/, tasks/ and train/, enough to run python -m train.finetune |
Run everything from the repo root with PYTHONPATH=src, because those data/ paths are
resolved relative to the working directory.
data/pairs.jsonl for the distilled student is not shipped. Building it needs a forward pass
over the whole train split to mine hard negatives, which is the slow step on a laptop and
cheap on a GPU, so molab_train.py builds it from distill/ on the box that will train on it.
| molab_train.py | a marimo notebook that pulls this repo and trains on CUDA |
The labels were corrected, and one earlier result was retracted
tool_select used to carry a boolean called multi_step, defined as two or more tool calls in the
Glaive conversation. The state handed to the model holds only the first user turn, and zero of
23,869 eligible rows carry two calls in that turn, while 7,610 are labelled true on the strength
of later ones. The label was not derivable from the input. A predictor given nothing but the name
of the selected tool scores AUROC 0.911 against the 0.921 that was published, so the model was
answering an easier question than the label claimed. That number is withdrawn.
needs_tool replaces it and asks whether the request needs a tool call at all, which the first
turn does determine. Haiku agrees with gold on it 98.7% of the time, against 72.4% on the label it
replaces. triage.priority was also collapsed from five rungs to the three the data uses, which
changed nothing: every backend, Haiku included, still sits at or below a constant predictor, and
that is published as a label-quality finding rather than fixed.
Teacher answers here were regenerated against the corrected questions. The superseded files are
not included: they carried zero needs_tool keys and 205 of their states had drifted into the
test and calib splits after the resampling.
- Downloads last month
- 97