Upload README.md with huggingface_hub
Browse files
README.md
CHANGED
|
@@ -196,7 +196,10 @@ spans 2.5β14.5 h while the 12 B-active 120β124 B MoEs finish in 1.8β3.4 h.
|
|
| 196 |
<details>
|
| 197 |
<summary><strong>Wall clock vs active parameters</strong> β tap to expand</summary>
|
| 198 |
|
| 199 |
-
<
|
|
|
|
|
|
|
|
|
|
| 200 |
|
| 201 |
</details>
|
| 202 |
|
|
@@ -212,7 +215,10 @@ fill) and a gray edge labelled **Abliterated**, because its Heretic abliteration
|
|
| 212 |
<details>
|
| 213 |
<summary><strong>Wall clock vs active parameters, perfect-score runs by country</strong> β tap to expand</summary>
|
| 214 |
|
| 215 |
-
<
|
|
|
|
|
|
|
|
|
|
| 216 |
|
| 217 |
</details>
|
| 218 |
|
|
@@ -222,7 +228,10 @@ finish in 1.75 h while a 4 B model takes 14.5 h.
|
|
| 222 |
<details>
|
| 223 |
<summary><strong>Wall clock vs total parameters</strong> β tap to expand</summary>
|
| 224 |
|
| 225 |
-
<
|
|
|
|
|
|
|
|
|
|
| 226 |
|
| 227 |
</details>
|
| 228 |
|
|
@@ -232,7 +241,10 @@ omitted); wall time still includes voided tasks.
|
|
| 232 |
<details>
|
| 233 |
<summary><strong>Wall clock vs total parameters, perfect-score runs by country</strong> β tap to expand</summary>
|
| 234 |
|
| 235 |
-
<
|
|
|
|
|
|
|
|
|
|
| 236 |
|
| 237 |
</details>
|
| 238 |
|
|
@@ -242,7 +254,10 @@ tokens take longer. The Spark outliers are the high-token runs, not the large mo
|
|
| 242 |
<details>
|
| 243 |
<summary><strong>Wall clock vs output tokens</strong> β tap to expand</summary>
|
| 244 |
|
| 245 |
-
<
|
|
|
|
|
|
|
|
|
|
| 246 |
|
| 247 |
</details>
|
| 248 |
|
|
@@ -253,7 +268,10 @@ outlier at ~18 k tokens/task.
|
|
| 253 |
<details>
|
| 254 |
<summary><strong>Verbosity vs active parameters</strong> β tap to expand</summary>
|
| 255 |
|
| 256 |
-
<
|
|
|
|
|
|
|
|
|
|
| 257 |
|
| 258 |
</details>
|
| 259 |
|
|
@@ -266,7 +284,10 @@ raw decode speed from the custom ternary kernels.
|
|
| 266 |
<details>
|
| 267 |
<summary><strong>Performance per byte</strong> β tap to expand</summary>
|
| 268 |
|
| 269 |
-
<
|
|
|
|
|
|
|
|
|
|
| 270 |
|
| 271 |
</details>
|
| 272 |
|
|
|
|
| 196 |
<details>
|
| 197 |
<summary><strong>Wall clock vs active parameters</strong> β tap to expand</summary>
|
| 198 |
|
| 199 |
+
<a href="plots/wall_vs_active.png" target="_blank">
|
| 200 |
+
<img src="plots/wall_vs_active.png" alt="Wall clock vs active parameters" style="width:100%; height:auto; max-width:100%;" />
|
| 201 |
+
</a>
|
| 202 |
+
<p><em>Tap chart to open full-resolution β pinch-zoom works in the image viewer.</em></p>
|
| 203 |
|
| 204 |
</details>
|
| 205 |
|
|
|
|
| 215 |
<details>
|
| 216 |
<summary><strong>Wall clock vs active parameters, perfect-score runs by country</strong> β tap to expand</summary>
|
| 217 |
|
| 218 |
+
<a href="plots/wall_vs_active_perfect_country.png" target="_blank">
|
| 219 |
+
<img src="plots/wall_vs_active_perfect_country.png" alt="Wall clock vs active parameters, perfect-score runs by country" style="width:100%; height:auto; max-width:100%;" />
|
| 220 |
+
</a>
|
| 221 |
+
<p><em>Tap chart to open full-resolution β pinch-zoom works in the image viewer.</em></p>
|
| 222 |
|
| 223 |
</details>
|
| 224 |
|
|
|
|
| 228 |
<details>
|
| 229 |
<summary><strong>Wall clock vs total parameters</strong> β tap to expand</summary>
|
| 230 |
|
| 231 |
+
<a href="plots/wall_vs_total.png" target="_blank">
|
| 232 |
+
<img src="plots/wall_vs_total.png" alt="Wall clock vs total parameters" style="width:100%; height:auto; max-width:100%;" />
|
| 233 |
+
</a>
|
| 234 |
+
<p><em>Tap chart to open full-resolution β pinch-zoom works in the image viewer.</em></p>
|
| 235 |
|
| 236 |
</details>
|
| 237 |
|
|
|
|
| 241 |
<details>
|
| 242 |
<summary><strong>Wall clock vs total parameters, perfect-score runs by country</strong> β tap to expand</summary>
|
| 243 |
|
| 244 |
+
<a href="plots/wall_vs_total_perfect_country.png" target="_blank">
|
| 245 |
+
<img src="plots/wall_vs_total_perfect_country.png" alt="Wall clock vs total parameters, perfect-score runs by country" style="width:100%; height:auto; max-width:100%;" />
|
| 246 |
+
</a>
|
| 247 |
+
<p><em>Tap chart to open full-resolution β pinch-zoom works in the image viewer.</em></p>
|
| 248 |
|
| 249 |
</details>
|
| 250 |
|
|
|
|
| 254 |
<details>
|
| 255 |
<summary><strong>Wall clock vs output tokens</strong> β tap to expand</summary>
|
| 256 |
|
| 257 |
+
<a href="plots/wall_vs_tokens.png" target="_blank">
|
| 258 |
+
<img src="plots/wall_vs_tokens.png" alt="Wall clock vs output tokens" style="width:100%; height:auto; max-width:100%;" />
|
| 259 |
+
</a>
|
| 260 |
+
<p><em>Tap chart to open full-resolution β pinch-zoom works in the image viewer.</em></p>
|
| 261 |
|
| 262 |
</details>
|
| 263 |
|
|
|
|
| 268 |
<details>
|
| 269 |
<summary><strong>Verbosity vs active parameters</strong> β tap to expand</summary>
|
| 270 |
|
| 271 |
+
<a href="plots/tokens_per_task_vs_active.png" target="_blank">
|
| 272 |
+
<img src="plots/tokens_per_task_vs_active.png" alt="Verbosity vs active parameters" style="width:100%; height:auto; max-width:100%;" />
|
| 273 |
+
</a>
|
| 274 |
+
<p><em>Tap chart to open full-resolution β pinch-zoom works in the image viewer.</em></p>
|
| 275 |
|
| 276 |
</details>
|
| 277 |
|
|
|
|
| 284 |
<details>
|
| 285 |
<summary><strong>Performance per byte</strong> β tap to expand</summary>
|
| 286 |
|
| 287 |
+
<a href="plots/perf_per_byte.png" target="_blank">
|
| 288 |
+
<img src="plots/perf_per_byte.png" alt="Performance per byte" style="width:100%; height:auto; max-width:100%;" />
|
| 289 |
+
</a>
|
| 290 |
+
<p><em>Tap chart to open full-resolution β pinch-zoom works in the image viewer.</em></p>
|
| 291 |
|
| 292 |
</details>
|
| 293 |
|