The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Error code: FileFormatMismatchBetweenSplitsError
Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
YAML Metadata Warning:empty or missing yaml metadata in repo card
Check out the documentation for more information.
NCP artifacts (v1)
datasets/ and rl/rl_prompts/ have had their prompt / prompt_full fields removed. Those
fields embedded the story-so-far, character sheets and chapter text for each section - verbatim
novel prose - which is not something to publish in a public repo, and it was 97% of the bytes.
Everything needed to rebuild them is here or already on the destination cluster:
- every row keeps
problem_id/question_id, which are exactly the idsncp_eval.data.example_id()emits (e.g.blue__chapter_0004__part_3_of_22); - the cohorts were built on the k8s cluster and live at
/mnt/disk/ncp_cohorts/{train,val,test}_examples.jsonl; - the prompt builder is in the
diversitygit repo (ncp_eval/stage_prompts.py,ncp_eval/v70_generate.py).
Rehydrate by joining on the id and re-rendering the prompt with the same builder the row was made
with. completion, metadata (incl. delta and threshold_pct), correct and source_method
are untouched, so the filtering and the targets are exactly as trained.
students/ holds the best-by-val-loss checkpoint per arm (Protocol B, the same resolution
ncp_merge_student.sh uses) plus trainer_state.json; see students/BEST_CHECKPOINTS.json.
Merged weights are rebuildable from base + adapter and are deliberately not here.
rl/rl_ground_truths/ is complete and unmodified - it holds control log-probs, no prose.
- Downloads last month
- 757