Focusing on efficiency for training high intelligence density models.
🥰 1660S my beloved
Chenhao Guan
Dsg2
AI & ML interests
"Built in a cave with a box of scraps"
Training small language models on low end consumer hardware.
Recent Activity
updated a collection about 3 hours ago
LS3 published a model about 3 hours ago
Dsg2/LS3-300M-A18M-Base new activity 1 day ago
AxiomicLabs/Open_SLM_Leaderboard:Add LS3-Micro-Dense-BaseOrganizations
None yet
Built on LS2's foundation, LS2.5 introduces a 16k vocab tokenizer to improve efficiency.
Experimental models: testing next gen algorithms
Focusing on efficiency for training high intelligence density models.
Experimental LaoShu models.
Built on LS2's foundation, LS2.5 introduces a 16k vocab tokenizer to improve efficiency.
LaoShu2 🐭 family: new architecture with engrams and top-2 routing, QAT without fp32 master weights.
Experimental models: testing next gen algorithms
LaoShu 🐭: tiny little models, no practicality but worth keeping anyways