Drop the load-time aside and the unrun page-set note
Browse files
README.md
CHANGED
|
@@ -61,7 +61,7 @@ once:
|
|
| 61 |
| | Build it yourself | This repository |
|
| 62 |
|---|---|---|
|
| 63 |
| Download | 9.0 GB (base + adapter) | 4.5 GB |
|
| 64 |
-
| Bring-up, base already on disk | about 1 minute and a 15 GB peak to merge and quantize; a bf16 load peaks at 16 GB |
|
| 65 |
| Memory while scoring | about 10 GB (bf16) | about 6.6 GB |
|
| 66 |
|
| 67 |
Measured on an M5 Pro with 24 GB, running macOS 27. The 8-bit process reached 8.5 GB during the longest suite.
|
|
@@ -115,8 +115,7 @@ counts 1,073 answers. The paired intervals are in [jaredpalmer/kev#162](https://
|
|
| 115 |
The local bf16 run matches the accuracy on Kev-4B's own model card for transfer-v4 (0.817) and decision-v7
|
| 116 |
(0.873), and is within 0.002 on hard-v1, devtools-v1 and documents-v1.
|
| 117 |
|
| 118 |
-
These are development splits. The locked test splits were not scored.
|
| 119 |
-
against the previous file (`485ace87`) and have not been rerun on this one.
|
| 120 |
|
| 121 |
## Use
|
| 122 |
|
|
|
|
| 61 |
| | Build it yourself | This repository |
|
| 62 |
|---|---|---|
|
| 63 |
| Download | 9.0 GB (base + adapter) | 4.5 GB |
|
| 64 |
+
| Bring-up, base already on disk | about 1 minute and a 15 GB peak to merge and quantize; a bf16 load peaks at 16 GB | about 4.8 GB |
|
| 65 |
| Memory while scoring | about 10 GB (bf16) | about 6.6 GB |
|
| 66 |
|
| 67 |
Measured on an M5 Pro with 24 GB, running macOS 27. The 8-bit process reached 8.5 GB during the longest suite.
|
|
|
|
| 115 |
The local bf16 run matches the accuracy on Kev-4B's own model card for transfer-v4 (0.817) and decision-v7
|
| 116 |
(0.873), and is within 0.002 on hard-v1, devtools-v1 and documents-v1.
|
| 117 |
|
| 118 |
+
These are development splits. The locked test splits were not scored.
|
|
|
|
| 119 |
|
| 120 |
## Use
|
| 121 |
|