Card: the new name, and the closed experiment's verdict
Browse files
README.md
CHANGED
|
@@ -2,7 +2,7 @@
|
|
| 2 |
license: mit
|
| 3 |
language:
|
| 4 |
- en
|
| 5 |
-
pretty_name: skycmd
|
| 6 |
task_categories:
|
| 7 |
- question-answering
|
| 8 |
tags:
|
|
@@ -14,12 +14,13 @@ configs:
|
|
| 14 |
data_files: data/test.parquet
|
| 15 |
---
|
| 16 |
|
| 17 |
-
# skycmd
|
| 18 |
|
| 19 |
-
The calibration and test questions of the LFM3-1.7B System 1 / System 2 router (`docs/lfm3-s1s2-prereg.md`): GSM8K and MATH, drawn with a fixed seed at pinned revisions.
|
| 20 |
|
| 21 |
Part of [skycmd](https://github.com/maxgfr/skycmd), tiny decision models trained from scratch in Rust that run in the
|
| 22 |
-
browser. Written by `scripts/lfm3-s1s2.sh data` at commit [`84c6d854`](https://github.com/maxgfr/skycmd/commit/84c6d854cdc2a9f6d951ebb55780060a01f38ca6)
|
|
|
|
| 23 |
Licences of every set: [DATA_LICENSES.md](https://github.com/maxgfr/skycmd/blob/main/DATA_LICENSES.md).
|
| 24 |
|
| 25 |
Each row is one question: `id`, `task` (`gsm8k` or `math`), `question` and `answer` (the final
|
|
@@ -44,7 +45,7 @@ Attribution goes to the upstream authors; the conversion is ours.
|
|
| 44 |
|
| 45 |
```python
|
| 46 |
from datasets import load_dataset
|
| 47 |
-
ds = load_dataset("maxgfr/skycmd-
|
| 48 |
```
|
| 49 |
|
| 50 |
skycmd wrote no row with a language model: the targets come from the upstream labels (some upstream
|
|
|
|
| 2 |
license: mit
|
| 3 |
language:
|
| 4 |
- en
|
| 5 |
+
pretty_name: skycmd math questions (GSM8K and MATH)
|
| 6 |
task_categories:
|
| 7 |
- question-answering
|
| 8 |
tags:
|
|
|
|
| 14 |
data_files: data/test.parquet
|
| 15 |
---
|
| 16 |
|
| 17 |
+
# skycmd math questions (GSM8K and MATH)
|
| 18 |
|
| 19 |
+
The calibration and test questions of the LFM3-1.7B System 1 / System 2 router (`docs/lfm3-s1s2-prereg.md`): GSM8K and MATH, drawn with a fixed seed at pinned revisions. That experiment is closed and failed its pre-registered gate: the router used 1,032 tokens per question on average, where the gate required at most half of the 958 used by always thinking (H3, compute halved). This data set was named `skycmd-lfm3-math` until 19/11/2026; the old address redirects here.
|
| 20 |
|
| 21 |
Part of [skycmd](https://github.com/maxgfr/skycmd), tiny decision models trained from scratch in Rust that run in the
|
| 22 |
+
browser. Written by `scripts/lfm3-s1s2.sh data` at commit [`84c6d854`](https://github.com/maxgfr/skycmd/commit/84c6d854cdc2a9f6d951ebb55780060a01f38ca6);
|
| 23 |
+
the script was removed from `main` on 19/11/2026 and is in the git history at that commit.
|
| 24 |
Licences of every set: [DATA_LICENSES.md](https://github.com/maxgfr/skycmd/blob/main/DATA_LICENSES.md).
|
| 25 |
|
| 26 |
Each row is one question: `id`, `task` (`gsm8k` or `math`), `question` and `answer` (the final
|
|
|
|
| 45 |
|
| 46 |
```python
|
| 47 |
from datasets import load_dataset
|
| 48 |
+
ds = load_dataset("maxgfr/skycmd-math", "calib")
|
| 49 |
```
|
| 50 |
|
| 51 |
skycmd wrote no row with a language model: the targets come from the upstream labels (some upstream
|