Skip to content

DeBERTa ablations and learning curve - #50

Merged
professor3333 merged 1 commit into
mainfrom
experiments/neural-ablations
Nov 9, 2026
Merged

professor3333 merged 1 commit into
mainfrom
experiments/neural-ablations

Conversation

@professor3333

Copy link
Copy Markdown
Owner

Scores the ablation-1 Kaggle run (15 runs) on the held-out half and the 10 hand-labelled dev docs; every seed reported, each variant paired with the same-seed full model. Learning curve 0.075/0.097/0.197 at 60/120/239 docs (reduced seeds all clearly worse); no-context and no-group not clearly different on held-out typed agreement. Adds scripts/gpu/score_ablations.sh; README, model card and GPU guide updated. Test split untouched.

Fifteen Kaggle runs of the run-3 recipe, scored on the held-out half and the 10 hand-labelled
dev documents with scripts/gpu/score_ablations.sh. DeBERTa is data-limited (typed agreement
0.075 / 0.097 / 0.197 at 60 / 120 / 239 documents, every reduced seed clearly worse); removing
context or the group objective makes no clear held-out difference. Every seed is reported.
@professor3333
professor3333 merged commit 2cf315e into main Nov 9, 2026
2 checks passed
@professor3333
professor3333 deleted the experiments/neural-ablations branch November 9, 2026 04:31
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