wbit5 commited on
Commit
14037f6
Β·
verified Β·
1 Parent(s): dedade8

publish from LatentPortPort c4a5ebf

Browse files
Files changed (1) hide show
  1. README.md +13 -0
README.md CHANGED
@@ -160,6 +160,19 @@ the two sizes differ in width, so the convolution state cannot be copied: the br
160
  [LFM Open License v1.0](https://huggingface.co/agentic-underground/lfm2-350m-to-1.2b-bridge/blob/main/LICENSE),
161
  which does not license Commercial Use by an entity with USD 10M or more in annual revenue.
162
 
 
 
 
 
 
 
 
 
 
 
 
 
 
163
  ### πŸ”€ A hybrid model β€” [`qwen3.5-0.8b-to-2b-bridge`](https://huggingface.co/agentic-underground/qwen3.5-0.8b-to-2b-bridge)
164
 
165
  Qwen3.5-0.8B reads; Qwen3.5-2B answers. Qwen3.5 mixes full-attention layers with Gated DeltaNet
 
160
  [LFM Open License v1.0](https://huggingface.co/agentic-underground/lfm2-350m-to-1.2b-bridge/blob/main/LICENSE),
161
  which does not license Commercial Use by an entity with USD 10M or more in annual revenue.
162
 
163
+ ### 🐍 Mamba-2, translated β€” [`granite-4.0-h-1b-to-h-tiny-bridge`](https://huggingface.co/agentic-underground/granite-4.0-h-1b-to-h-tiny-bridge)
164
+
165
+ Granite 4.0 H 1B reads; H Tiny, a 64-expert mixture, answers. Mamba-2 state cannot be copied between
166
+ separately trained sizes, so the bridge translates each recurrent layer's **inputs**, and H Tiny
167
+ builds its own state from them.
168
+
169
+ | | |
170
+ |:--|:--|
171
+ | 🎯 **Held-out questions** | **7 of 7**, the same as each model alone |
172
+ | πŸ” **Fidelity (NCR)** | **0.924** with the paper's learned correction (0.916 from the maps alone) |
173
+ | πŸ§ͺ **Control (wrong donor)** | βˆ’0.175 |
174
+ | ⚑ **Time to first token** | **about 2Γ—** sooner on one RTX 3090 at ~256 tokens (1.38Γ— to 2.06Γ— across three runs; its experts route on the host, so it varies with load) |
175
+
176
  ### πŸ”€ A hybrid model β€” [`qwen3.5-0.8b-to-2b-bridge`](https://huggingface.co/agentic-underground/qwen3.5-0.8b-to-2b-bridge)
177
 
178
  Qwen3.5-0.8B reads; Qwen3.5-2B answers. Qwen3.5 mixes full-attention layers with Gated DeltaNet