Alisa Liu
@alisawuffles.bsky.social did:plc:usfej5yc6glifzm5d73ddhvk
phd student at @uwcse
Bluesky’a katılma tarihi:
- Takipçi
- 387
- Takip edilen
- 197
- Post
- 12
- Sıralamaya girdiği gün
- 0
Özet
Alisa Liu (@alisawuffles.bsky.social), 19 Kasım 2024 tarihinde katılan bir Bluesky kullanıcısı. 387 takipçisi var, 197 hesabı takip ediyor ve 12 post paylaştı.
Sıralama geçmişi
Bu hesabın postlarının Debya günlük listesine girdiği en iyi günler.
Bu hesap henüz günlük sıralamaya girmedi.
Son beğendiği postlar
Bluesky ağındaki herkese açık beğeniler.
3. the default dynamics of the alternative that you’re proposing are rich-get-richer with limited incentive for quality. i think you can see this quality issue in the not-for-peer-review reports that, ahem, some of the big LLM labs put out
Postu göster1. from a macro perspective (progress of the field, not individual decisions), peer review has a great track record 2. it’s a somewhat fair and noisy signal of quality that can draw attention to papers that wouldn’t otherwise get read +
Postu gösterDo you get idioms? Because one would think that is where this is mostly needed ("kicked the bucket"), but BPE usually works based on raw counts, so "kicked the bucket" might not win the frequency game. Right?
Postu gösterVery cool, I'd love to plug SuperBPE into LMs and see how it does!
Postu gösterHumans store thousands of multi-word expressions like "of course" in their mental lexicon, but current tokenizers don't support multi-word tokens. Enter SuperBPE, a tokenizer that lifts this restriction and brings substantial gains in efficiency and performance! 🚀 Details 👇
Postu göster
Blokladığı hesaplar
Bluesky’da bloklar herkese açık kayıtlardır.
