Skip to content

V65: an unlabelled file becomes a submission - #165

Merged
dapiced merged 1 commit into
mainfrom
dom/modest-darwin-4oe9q0
Nov 13, 2026
Merged

dapiced merged 1 commit into
mainfrom
dom/modest-darwin-4oe9q0

Conversation

@dapiced

@dapiced dapiced commented Nov 13, 2026

Copy link
Copy Markdown
Owner

What

The second part of F8, the competition path. Under a scored batch, a new Submission file block writes what a platform expects: an identifier and a prediction, under the names the user gives.

  • train/submission.ts (pure): three forms of prediction (class by name, probability of the second class in binary, one probability per class in multiclass, value in regression). Every row of the scored file is kept in its order; a missing or repeated identifier is written and counted, never dropped.
  • train/refit.ts (pure): the inspected family fitted again on every labelled row, same columns, seed, caps and class weighting. Used for the submission only; it enters neither the leaderboard, nor the export, nor the saved run.
  • Worker request build-submission on the last scored file, which stays in the worker. The worker keeps the configuration its artifacts were trained with, so a refit cannot repeat another run.
  • SubmissionPanel inside BatchScorePanel, strings in French and English.

Measured before building

Does fitting again on every row pay? Six demos, ten splits each, a quarter of the rows kept away from both models, every family:

Demo Labelled rows Mean gain on the rows kept away Better / worse
reviews 180 +5.0 accuracy points 51 / 11 of 70
mpg 298 RMSE down 3 % 47 / 13 of 60
energy 180 RMSE down 2 % 44 / 16 of 60
titanic 668 +0.8 accuracy point 41 / 20 of 70
fraud 480 none 21 / 25 of 70
iris 112 none 14 / 17 of 70

The plan expected a gain « clearer on small files ». That was wrong: iris is the smallest demo and gains nothing. The gain follows the learning curve, not the size. So the refit is offered, never the default, and the result line names the model that predicted: the measured one, or its sibling nobody measured.

Design rules

  • The lab guesses neither the identifier nor the names. It points at a column the training file read as an identifier when there is exactly one, and does not choose it.
  • A submission is not a score: no metric of the platform, no private rows, no rank promised. A public leaderboard wears like a test set (V60).
  • The blend cannot be refitted (its members are picked on validation rows): the option is disabled with the reason.

Not in this wave

Sending anything to a platform, measuring the refitted model, a submission from a re-imported model, applying the decision threshold to the written label, blending several models into one submission. The glass box stays proposed.

Docs

Score a batch (new section), limits, ML reference and refusals in both languages, README, PLAN row V65, changelog, version 1.65.0.

Checks

  • Lint, typecheck (apart from the xlsx types unreachable in my sandbox), format: clean.
  • Unit: 1246 passed locally, 5 failed, all in spreadsheet.test.ts (the xlsx tarball cannot be fetched in my sandbox; CI is the reference). 27 new tests.
  • E2E: 131 passed locally, 1 failed (lab.spec.ts:31, the Excel upload, same cause). New e2e/submission.spec.ts scores a file whose identifiers are out of order and checks the order survives, then builds the refitted probabilities.
  • Screenshot review of the panel in French, light desktop and dark mobile; a sideways overflow on mobile was found there and fixed.

Generated by Claude Code

Under a scored batch, a block writes the file a competition platform
expects: an identifier and a prediction under the names the user gives,
every row kept in its order, missing and repeated identifiers counted.
The lab guesses neither the identifier nor the names.

The same family can be fitted again on every labelled row. Measured
first, on the six demos against rows kept away from both models: +5.0
accuracy points on reviews, +0.8 on titanic, 2 to 3 % of RMSE on mpg
and energy, nothing on fraud and iris. The gain follows the learning
curve, not the size of the file, so the refit is offered, never the
default, and announced as a model nobody measured.

Docs (score a batch, limits, reference, refusals) in both languages,
README, PLAN row V65, version 1.65.0.

Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01AyBsUet8WybTG986ubroar
@dapiced
dapiced marked this pull request as ready for review November 13, 2026 17:43
@dapiced
dapiced merged commit b096ac5 into main Nov 13, 2026
7 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.

2 participants