Spaces:
Running
Running
publish from LatentPortPort 473f8ef
Browse files
README.md
CHANGED
|
@@ -83,6 +83,21 @@ Gemma 3 270M reads; Gemma 3 1B answers, across sliding-window and global attenti
|
|
| 83 |
βοΈ **Gated, and not Apache-2.0:** it is a Gemma Model Derivative, distributed under the
|
| 84 |
[Gemma Terms of Use](https://ai.google.dev/gemma/terms).
|
| 85 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 86 |
### π A hybrid model β [`qwen3.5-0.8b-to-2b-bridge`](https://huggingface.co/agentic-underground/qwen3.5-0.8b-to-2b-bridge)
|
| 87 |
|
| 88 |
Qwen3.5-0.8B reads; Qwen3.5-2B answers. Qwen3.5 mixes full-attention layers with Gated DeltaNet
|
|
@@ -124,7 +139,8 @@ and measured.
|
|
| 124 |
- π **Attributed** β PrattleNot is a derivative work of [LatentPort](https://latentport.com/) by
|
| 125 |
Simon Villani ([source](https://github.com/LatentPort/latentport), Apache-2.0). Our bridges are
|
| 126 |
Apache-2.0, except those fitted from Gemma, which are Gemma Model Derivatives, gated and distributed
|
| 127 |
-
under the [Gemma Terms of Use](https://ai.google.dev/gemma/terms)
|
|
|
|
| 128 |
|
| 129 |
## Elsewhere
|
| 130 |
|
|
|
|
| 83 |
βοΈ **Gated, and not Apache-2.0:** it is a Gemma Model Derivative, distributed under the
|
| 84 |
[Gemma Terms of Use](https://ai.google.dev/gemma/terms).
|
| 85 |
|
| 86 |
+
### π¦ A fourth family β [`llama-3.2-1b-to-3b-bridge`](https://huggingface.co/agentic-underground/llama-3.2-1b-to-3b-bridge)
|
| 87 |
+
|
| 88 |
+
**Built with Llama.** Llama 3.2 1B reads; Llama 3.2 3B answers, across different head widths (64
|
| 89 |
+
and 128).
|
| 90 |
+
|
| 91 |
+
| | |
|
| 92 |
+
|:--|:--|
|
| 93 |
+
| π **Fidelity (NCR)** | **0.965** with the paper's learned correction (0.818 from the maps alone), our highest yet |
|
| 94 |
+
| π§ͺ **Control (wrong donor)** | β0.040 |
|
| 95 |
+
| π― **Held-out questions** | **6 of 7**, the same as the 3B alone (the 1B: 7). No answer gain yet |
|
| 96 |
+
| β‘ **Time to first token** | **1.70Γ** sooner (95% CI 1.66β1.74) on one RTX 3090 at ~256 tokens; ceiling 2.54Γ |
|
| 97 |
+
|
| 98 |
+
βοΈ **Gated, and not Apache-2.0:** it is a derivative work of Llama 3.2, distributed under the
|
| 99 |
+
[Llama 3.2 Community License](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct/blob/main/LICENSE.txt).
|
| 100 |
+
|
| 101 |
### π A hybrid model β [`qwen3.5-0.8b-to-2b-bridge`](https://huggingface.co/agentic-underground/qwen3.5-0.8b-to-2b-bridge)
|
| 102 |
|
| 103 |
Qwen3.5-0.8B reads; Qwen3.5-2B answers. Qwen3.5 mixes full-attention layers with Gated DeltaNet
|
|
|
|
| 139 |
- π **Attributed** β PrattleNot is a derivative work of [LatentPort](https://latentport.com/) by
|
| 140 |
Simon Villani ([source](https://github.com/LatentPort/latentport), Apache-2.0). Our bridges are
|
| 141 |
Apache-2.0, except those fitted from Gemma, which are Gemma Model Derivatives, gated and distributed
|
| 142 |
+
under the [Gemma Terms of Use](https://ai.google.dev/gemma/terms), and those fitted from Llama, which
|
| 143 |
+
are Built with Llama, gated and distributed under the Llama 3.2 Community License.
|
| 144 |
|
| 145 |
## Elsewhere
|
| 146 |
|