Add ideas from GitHub issues labeled idea - 2026-10-27

Merged
YA
15d ago
+55-0

Migrates the 12 open GitHub issues labeled idea into the topic IDEAS.md files.

Per file

  • model_design/IDEAS.md: #42, #71, #72 (model failure analysis and explanation from embeddings/clusters, next to the existing "Make model failure analysis from embedded clustering…" and "analyse model failure from closest in training set" entries); #63 (hyperparameter choice from similar datasets, next to "Is the best model+config for dataset A the same for dataset B…").
  • data_selection/IDEAS.md: #75 (how much data a task needs depending on training status); #62 (selecting samples for test-time in-context injection).
  • out_of_distribution/IDEAS.md: #70 (OOD detection by comparing predictions inside a semantic cluster).
  • drift/IDEAS.md: #64 (clustering to detect drift from embedding distributions).
  • losses/IDEAS.md: #60 (loss weighting from full-training-set embeddings).
  • batching/IDEAS.md: #55 (batch differences to explain gradient variations).
  • IDEAS.md (root): #74, #58 — see below.

Possible new topics

  • #74 "Is reidentification possible from simple embedding comparison ?" — suggested topic: retrieval (similarity search / re-identification). Not data selection, drift, OOD or labeling (not about choosing, monitoring or annotating data); not model design (no architecture/config search).
  • #58 "make data registration/alignement based on foundational features" — suggested topic: preprocessing (registration, alignment, normalization of raw data). Not augmentation (no synthetic variants), not data selection/labeling/drift/OOD; other topics are about training or monitoring.

Duplicates

None.

In progress (covered by an open PR)

None.

🤖 Generated with Claude Code

Summary by cubic

Migrates 12 open GitHub issues labeled idea into the topic-specific IDEAS.md files.

  • Adds entries to model_design, data_selection, out_of_distribution, drift, losses, batching, and root IDEAS.md.
  • Places issues about embedding-based failure analysis and dataset similarity in model_design, and adds two new topics (retrieval, preprocessing) to the root file for issues not covered by existing files.
  • No duplicates or in-progress work found among the migrated issues.