Tiny language models that write cheerful stories by day and spooky ones at night, triggered only by a time header.
Henry PRO
hbfreed
AI & ML interests
None yet
Recent Activity
updated a dataset 8 days ago
hbfreed/tinydark-gram-data published a dataset 8 days ago
hbfreed/tinydark-gram-data updated a model 28 days ago
hbfreed/tiny-dark-stories-bigOrganizations
Winnow
Variable-width MoE expert pruning (Winnow): pruned OLMoE, Qwen3.6-35B-A3B and poolside Laguna-S-2.1 checkpoints, unhealed and healed.
-
hbfreed/Laguna-S-2.1-winnow-keep75
Text Generation • 89B • Updated • 8 -
hbfreed/Laguna-S-2.1-winnow-keep50
Text Generation • 61B • Updated • 10 -
hbfreed/Qwen3.6-35B-A3B-winnow-keep75
Text Generation • 27B • Updated • 20 • 1 -
hbfreed/Qwen3.6-35B-A3B-winnow-keep50
Text Generation • 19B • Updated • 16 • 1
Tiny Dark Stories
Tiny language models that write cheerful stories by day and spooky ones at night, triggered only by a time header.
Beyond Owls: on-policy distillation
Backdoor transfer through on-policy distillation and self-distilled teachers on Qwen3.5-9B. Code: github.com/hbfreed/on-policy-subliminal-learning
Winnow
Variable-width MoE expert pruning (Winnow): pruned OLMoE, Qwen3.6-35B-A3B and poolside Laguna-S-2.1 checkpoints, unhealed and healed.
-
hbfreed/Laguna-S-2.1-winnow-keep75
Text Generation • 89B • Updated • 8 -
hbfreed/Laguna-S-2.1-winnow-keep50
Text Generation • 61B • Updated • 10 -
hbfreed/Qwen3.6-35B-A3B-winnow-keep75
Text Generation • 27B • Updated • 20 • 1 -
hbfreed/Qwen3.6-35B-A3B-winnow-keep50
Text Generation • 19B • Updated • 16 • 1
Variable FlexOlmo
Pruned and distilled variants of Flex-math-2x7B-1T with variable-width expert MLPs.