Skip to content

docs(EXPERIMENTS): E1 corrected-artifact sweep verdict — heb outlier 9.34% -> 0.29% flips - #111

Merged
ronaldtse merged 2 commits into
mainfrom
docs/e1-sweep-verdict
Aug 31, 2026
Merged

docs(EXPERIMENTS): E1 corrected-artifact sweep verdict — heb outlier 9.34% -> 0.29% flips#111
ronaldtse merged 2 commits into
mainfrom
docs/e1-sweep-verdict

Conversation

@ronaldtse

Copy link
Copy Markdown
Contributor

Completes E1's arc: all four shipped int8 artifacts rebuilt body-int8/head-fp32 and gated through the full stack (parity in-zip, margins, confident-flip budget). The heb-diac-1.1 outlier that motivated the investigation lands at 0.29 percent flips (from 9.34 percent, 32x) with ~zero confident flips at +0.4 percent size. Includes the two infra fixes found en route (refresh_member_shas PR #99; double-split entrypoint PR #109) and records the swap-in options as an owner decision (in-place + index-v2 vs parallel -int8-head32 ids). tha-g2p-small in flight.

Ronald Tse added 2 commits August 31, 2026 18:59
…ts again

'AttributeError: list object has no attribute split' on every plain
'modal run src/gpu/modal_export.py --model X' — the un-suffixed
entrypoint has been broken since export_model learned to split;
parity/margins/rebuild_int8 paths were unaffected.
…9.34% flips -> 0.29% (32x) with head-fp32

All four shipped int8 artifacts rebuilt body-int8/head-fp32 and gated
through the full stack (parity in-zip, margins, confident-flip budget
<=1%). The outlier that motivated E1 — heb-diac-1.1 int8 at 9.34%
flips, 80% at confident positions — lands at 0.29% flips and ~zero
confident flips at +0.4% size. Swap-in options (in-place + index-v2 vs
parallel -int8-head32 ids) recorded as an owner decision; tha-g2p-small
in flight.
@ronaldtse
ronaldtse merged commit 9e6b5df into main Aug 31, 2026
5 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant