Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -20,8 +20,14 @@ blob they were computed for.
|
|
| 20 |
|
| 21 |
| file | bytes | contents |
|
| 22 |
| --- | --- | --- |
|
| 23 |
-
| `
|
| 24 |
-
| `
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 25 |
|
| 26 |
The decode graph is compiled for a fixed 256-slot KV cache and emits both logits
|
| 27 |
and all 13 residual stream taps, so both lenses read from one forward pass.
|
|
|
|
| 20 |
|
| 21 |
| file | bytes | contents |
|
| 22 |
| --- | --- | --- |
|
| 23 |
+
| `weights-*.bin` | 495,399,936 | transformer weights, fp32, laid out for the exported graph |
|
| 24 |
+
| `lens-*.bin` | 30,670,848 | fitted J-lens Jacobians, 13 layers |
|
| 25 |
+
|
| 26 |
+
Weights are delivered as 16 MiB content-addressed chunks. Each file name carries
|
| 27 |
+
the SHA-256 of its own contents, so a chunk URL always refers to exactly one
|
| 28 |
+
sequence of bytes: safe to cache indefinitely, and impossible to pair with a
|
| 29 |
+
manifest describing a different packing. The manifests listing them live in the
|
| 30 |
+
site repository beside the loader.
|
| 31 |
|
| 32 |
The decode graph is compiled for a fixed 256-slot KV cache and emits both logits
|
| 33 |
and all 13 residual stream taps, so both lenses read from one forward pass.
|