Spaces:
Running
Running
publish from LatentPortPort f9441a9
Browse files
README.md
CHANGED
|
@@ -68,8 +68,7 @@ SmolLM2-360M reads; SmolLM2-1.7B, about five times larger and shaped differently
|
|
| 68 |
| π― **Held-out questions** | **7 of 7**, as the 1.7B reading everything (the 360M alone: 6) |
|
| 69 |
| π **Fidelity (NCR)** | **0.845** with the paper's learned correction (0.611 from the maps alone) |
|
| 70 |
| π§ͺ **Control (wrong donor)** | β0.266 |
|
| 71 |
-
|
| 72 |
-
It runs server-side for now; its speed-up is not measured yet, so the card claims none.
|
| 73 |
|
| 74 |
### 𧬠A second family β [`qwen3-0.6b-to-1.7b-bridge`](https://huggingface.co/agentic-underground/qwen3-0.6b-to-1.7b-bridge)
|
| 75 |
|
|
@@ -81,8 +80,7 @@ carries beyond one family, fitted with the same tools and measured on held-out q
|
|
| 81 |
| π **Fidelity (NCR)** | **0.805** with the paper's learned correction (0.645 from the maps alone) |
|
| 82 |
| π― **Held-out questions** | **7 of 7** correct, as many as the 1.7B reading everything |
|
| 83 |
| π§ͺ **Control (wrong donor)** | β0.127: the bridge fed another question's memory scores far lower |
|
| 84 |
-
|
| 85 |
-
It runs server-side for now; its speed-up is not measured yet, so the card claims none.
|
| 86 |
|
| 87 |
Each new pair gets its own repository here and joins the collection once its weights are fitted
|
| 88 |
and measured.
|
|
|
|
| 68 |
| π― **Held-out questions** | **7 of 7**, as the 1.7B reading everything (the 360M alone: 6) |
|
| 69 |
| π **Fidelity (NCR)** | **0.845** with the paper's learned correction (0.611 from the maps alone) |
|
| 70 |
| π§ͺ **Control (wrong donor)** | β0.266 |
|
| 71 |
+
| β‘ **Time to first token** | **1.67Γ** sooner (95% CI 1.66β1.69) on two RTX 3090s at ~256 tokens; ceiling 2.59Γ |
|
|
|
|
| 72 |
|
| 73 |
### 𧬠A second family β [`qwen3-0.6b-to-1.7b-bridge`](https://huggingface.co/agentic-underground/qwen3-0.6b-to-1.7b-bridge)
|
| 74 |
|
|
|
|
| 80 |
| π **Fidelity (NCR)** | **0.805** with the paper's learned correction (0.645 from the maps alone) |
|
| 81 |
| π― **Held-out questions** | **7 of 7** correct, as many as the 1.7B reading everything |
|
| 82 |
| π§ͺ **Control (wrong donor)** | β0.127: the bridge fed another question's memory scores far lower |
|
| 83 |
+
| β‘ **Time to first token** | **1.47Γ** sooner (95% CI 1.46β1.48) on two RTX 3090s at ~256 tokens; ceiling 2.03Γ |
|
|
|
|
| 84 |
|
| 85 |
Each new pair gets its own repository here and joins the collection once its weights are fitted
|
| 86 |
and measured.
|