Spaces:
Sleeping
Sleeping
Open external links in a new tab
Browse files- static/index.html +1 -1
static/index.html
CHANGED
|
@@ -82,7 +82,7 @@ h2{font-size:14px;font-weight:600;margin:0 0 12px;color:var(--muted)}
|
|
| 82 |
<section class="about">
|
| 83 |
<h2>About the model</h2>
|
| 84 |
<p>TypeSafe AI's <a href="https://docs.typesafe.ai/introduction" target="_blank" rel="noopener">Jev</a> has got people excited about "System One" models: models that don't generate text, but read a state and return typed answers with probabilities that software can use directly. This page does the same thing for one narrow job. The state is a dataset's columns and first row. The answer is a set of task tags, each with a probability. The model is small enough to fine-tune yourself on Jobs and to run on a free CPU.</p>
|
| 85 |
-
<p>The tags come from <a href="https://huggingface.co/davanstrien/hub-task-tagger-gliner2.5-base">GLiNER2.5-base</a>, a 194M-parameter model from <a href="https://huggingface.co/fastino/gliner2.5-base-v1">Fastino</a> that reads a text and a list of labels together. It was fine-tuned on 16,000 Hub datasets and their owners' tags, in 17 minutes on one GPU with <a href="https://huggingface.co/datasets/uv-scripts/classification/blob/main/train-gliner2.py">this training script</a>.</p>
|
| 86 |
<p>On 3,000 newer datasets from owners it never saw, its first suggestion matches an owner's tag 69% of the time. That is a noisy target: owners often leave out tags that fit, and in a hand-checked sample about one dataset in ten was missing one. A suggestion that isn't in the owner's tags is not always wrong.</p>
|
| 87 |
|
| 88 |
</section>
|
|
|
|
| 82 |
<section class="about">
|
| 83 |
<h2>About the model</h2>
|
| 84 |
<p>TypeSafe AI's <a href="https://docs.typesafe.ai/introduction" target="_blank" rel="noopener">Jev</a> has got people excited about "System One" models: models that don't generate text, but read a state and return typed answers with probabilities that software can use directly. This page does the same thing for one narrow job. The state is a dataset's columns and first row. The answer is a set of task tags, each with a probability. The model is small enough to fine-tune yourself on Jobs and to run on a free CPU.</p>
|
| 85 |
+
<p>The tags come from <a href="https://huggingface.co/davanstrien/hub-task-tagger-gliner2.5-base" target="_blank" rel="noopener">GLiNER2.5-base</a>, a 194M-parameter model from <a href="https://huggingface.co/fastino/gliner2.5-base-v1" target="_blank" rel="noopener">Fastino</a> that reads a text and a list of labels together. It was fine-tuned on 16,000 Hub datasets and their owners' tags, in 17 minutes on one GPU with <a href="https://huggingface.co/datasets/uv-scripts/classification/blob/main/train-gliner2.py" target="_blank" rel="noopener">this training script</a>.</p>
|
| 86 |
<p>On 3,000 newer datasets from owners it never saw, its first suggestion matches an owner's tag 69% of the time. That is a noisy target: owners often leave out tags that fit, and in a hand-checked sample about one dataset in ten was missing one. A suggestion that isn't in the owner's tags is not always wrong.</p>
|
| 87 |
|
| 88 |
</section>
|