Repository navigation
Conversation
…tency bench
- eval_plane_k: subset K in {1,2,4,8} sweep over best/ (subset
labeled, never a gate number; full-set at best K optional)
- export_plane_onnx: one graph (input_ids, plane_ids) -> class
logits, host-side K-pass loop (no KV cache); dynamic int8; CPU
ms/window vs the shipped seq2seq int8static tier
* origin/main: feat(run-017): plane-factorized diacritization - Stoicheia WO entered feat(r9): YallaMorph/CamelMorph morph aux teacher + RIDE direction probe
Grid evidence (subset): quality monotonically improving through the last checkpoint (3.30/2.99/2.93 - NOT overfit, undertrained), K flat across 2/4/8 (K=2 wins latency: 1145ms vs shipped 3459ms).
… epoch, K=2 - SOTA gate 2.2864
…lation run-019 measured 3.0078 at 1 epoch (beats Gemini-Flash, gate open). Arm A: EPOCHS=2 resume (epoch effect measured -1.17pp at small tier). Arm B: ByT5-small plane + r7 teacher labels x6, 3 epochs - the supervision lever, single-variable vs run-018's raw text.
The first E2 eval reused epoch-1's eval_progress.jsonl: every window was already 'saved', so it regenerated nothing and re-scored epoch-1 predictions (DER bit-identical at 3.0078). EVAL_DONE_E2B + eval_progress_EVAL_DONE_E2B.jsonl force full regeneration.
…d exporter build_plane_zip emits the secryst-py PlaneModel.from_zip contract (metadata kind=plane + member sha256s, plane.onnx, classes.json); exporter gains --checkpoint/--backbone/--prefix/--bench-skip.
The Arabic run-018 recipe applied to Hebrew over the v4 combined corpus; eval on nakdimon test with the exact seq2seq_der protocol (heb-diac-1.1's 16.44 DER gate). Per-marker eval progress file (the run-019 E2 lesson applied by design).
The first run-021 verdict read DER 32.64% from whitespace collapse in stitching (double spaces/newlines shifted all subsequent units); re-derived from cached window predictions with exact separator re-attachment: DER 12.48% — the 16.44 gate is cleared.
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Follow-on levers for the plane-factorized run, ready to fire when best/ lands: K in {1,2,4,8} sweep (subset labeled, full-set optional) and the ONNX export + dynamic int8 + CPU ms/window bench (one graph, host-side K-pass loop, no KV cache).