Rewrite dataset attribution storage + 3 metrics#413

Open

ocg-goodfire wants to merge 17 commits intodevfrom

feature/faster-dataset-attributions-2

Commits on Feb 23, 2026

add worktrees to ignore
ocg-goodfire
committed
Rewrite dataset attribution storage: dict-of-dicts, canonical names, 3 metrics

ocg-goodfire
and
claude
committed
Fix alive_targets iteration: use torch.where for indices, not bool tolist

ocg-goodfire
and
claude
committed
Fix KeyError for embed source: CI dict doesn't include embedding layer

ocg-goodfire
and
claude
committed
Fix scatter_add OOB: use embedding num_embeddings instead of tokenizer vocab_size

ocg-goodfire
and
claude
committed
Split run.py into run_worker.py and run_merge.py

ocg-goodfire
and
claude
committed
Correct attr_abs via backprop through |target|, reorganise method signatures

ocg-goodfire
and
claude
committed
Add merge_mem config (default 200G) to prevent merge OOM

ocg-goodfire
and
claude
committed
Add 3-metric selection to dataset attributions in app

ocg-goodfire
and
claude
committed
Allow bare s-prefixed run IDs everywhere (e.g. "s-17805b61")

ocg-goodfire
and
claude
committed
Fix AttributionRepo.open skipping valid subruns due to old-format dirs

ocg-goodfire
and
claude
committed
Fix 3s lag on attribution metric toggle: O(V) linear scan per pill

ocg-goodfire
and
claude
committed
Ship token strings from backend instead of resolving vocab IDs in frontend

ocg-goodfire
and
claude
committed
Hide negative attribution column for non-signed metrics

ocg-goodfire
and
claude
committed
Narrow frontend types: SignedAttributions vs UnsignedAttributions

ocg-goodfire
and
claude
committed
Update dataset_attributions CLAUDE.md for new storage format and 3 metrics

ocg-goodfire
and
claude
committed
Remove stray worktree files from PR

ocg-goodfire
and
claude
committed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Rewrite dataset attribution storage + 3 metrics#413

Rewrite dataset attribution storage + 3 metrics#413
ocg-goodfire wants to merge 17 commits intodevfrom
feature/faster-dataset-attributions-2

Commits on Feb 23, 2026

add worktrees to ignore

Rewrite dataset attribution storage: dict-of-dicts, canonical names, 3 metrics

Fix alive_targets iteration: use torch.where for indices, not bool tolist

Fix KeyError for embed source: CI dict doesn't include embedding layer

Fix scatter_add OOB: use embedding num_embeddings instead of tokenizer vocab_size

Split run.py into run_worker.py and run_merge.py

Correct attr_abs via backprop through |target|, reorganise method signatures

Add merge_mem config (default 200G) to prevent merge OOM

Add 3-metric selection to dataset attributions in app

Allow bare s-prefixed run IDs everywhere (e.g. "s-17805b61")

Fix AttributionRepo.open skipping valid subruns due to old-format dirs

Fix 3s lag on attribution metric toggle: O(V) linear scan per pill

Ship token strings from backend instead of resolving vocab IDs in frontend

Hide negative attribution column for non-signed metrics

Narrow frontend types: SignedAttributions vs UnsignedAttributions

Update dataset_attributions CLAUDE.md for new storage format and 3 metrics

Remove stray worktree files from PR