MELT Workshop

meltworkshop.bsky.social

10 October 2025 15:53

Keynote talk: Apertus: Democratizing Open and Compliant LLMs for Global Language Environments. Imanol Schlag introduces Apertus, a fully open suite of LLMs with a focus on compliance, transparency, and multilingual representation training across 1000+ languages. 🌍🤖

Deniz Bayazit

bayazitdeniz.bsky.social

26 September 2025 01:18

5/ Looking closer, feature sharing has limits: in Hindi & Arabic, overlap stays low even at 341B tokens. This may be due to richer agreement systems (e.g., verbs agreeing w/ subjects & objects) forcing BLOOM to keep language-specific features—or simply data scarcity!

Deniz Bayazit

bayazitdeniz.bsky.social

26 September 2025 01:18

4/ In #multilingual models, cross-language feature overlap starts low and rises with training. At 6B tokens in BLOOM, most detectors are language-specific or for punctuation; by 341B tokens shared crosslingual features emerge, capturing syntactic abstractions over token patterns.

Deniz Bayazit

bayazitdeniz.bsky.social

26 September 2025 01:18

3/ Which features matter early but fade, and which gain importance later? In Pythia, token-level detectors drop out, while higher-level grammatical features—like plural-noun detectors and nouns formed from verbs (e.g., runner from run)—strengthen by 286B tokens.

Deniz Bayazit

bayazitdeniz.bsky.social

26 September 2025 01:18

2/ We align critical checkpoints for a task with sparse crosscoders, measure each feature’s causal role, and introduce RelIE to compare their influence across checkpoints. This lets us trace how internal features shift—and when they matter—in models like Pythia, OLMo, and BLOOM.