Spaces:
Running
Running
Ctrl+K
Update logbook: Repro: Do Transformers Need Three Projections? Systematic Study of QKV Variants
5f2a7a1 verified - bug-report
- c1-absolute-perplexity-deflation
- c1-c2-training-baseline-vs-q-k-v-scaled-down-cpu
- c2-relative-gap-dilution-headline-falsification
- c3-pad-fraction-measurement
- claim-2-q-k-v-fully-shared-variant
- claim-3-q-k-v-mqa-gqa-4
- claim-5-k-v-cosine-similarity
- corrected-claim-statement
- limitations
- metric-provenance-tables-4-9-vs-table-11
- summary
- 958 Bytes