Spaces:
Runtime error
Runtime error
JoFrost
commited on
Commit
·
8330b1d
1
Parent(s):
155a7fa
feat: add text
Browse files
app.py
CHANGED
@@ -24,6 +24,44 @@ params = {
|
|
24 |
"filename" : "layer_31/width_16k/average_l0_76/params.npz"
|
25 |
}
|
26 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
27 |
C = 0.01
|
28 |
|
29 |
model_name = params["model_name"]
|
@@ -168,58 +206,61 @@ def get_feature_iframe(feature):
|
|
168 |
html = gr.HTML(html_content)
|
169 |
return html
|
170 |
|
171 |
-
with gr.Blocks() as demo:
|
172 |
-
with gr.
|
173 |
-
with gr.
|
174 |
-
|
175 |
-
|
176 |
-
|
177 |
-
|
178 |
-
|
179 |
-
|
180 |
-
|
181 |
-
|
182 |
-
|
183 |
-
|
184 |
-
|
185 |
-
|
186 |
-
|
187 |
-
|
188 |
-
|
189 |
-
|
190 |
-
|
191 |
-
|
192 |
-
|
193 |
-
|
194 |
-
|
195 |
-
|
196 |
-
|
197 |
-
|
198 |
-
|
199 |
-
|
200 |
-
|
201 |
-
|
202 |
-
|
203 |
-
|
204 |
-
|
205 |
-
|
206 |
-
|
207 |
-
|
208 |
-
|
209 |
-
).then(
|
210 |
-
fn=get_feature_iframe,
|
211 |
-
inputs=[dropdown],
|
212 |
-
outputs=[html]
|
213 |
-
)
|
214 |
-
|
215 |
-
dropdown.change(
|
216 |
fn=get_highlighted_text,
|
217 |
-
|
218 |
-
|
219 |
-
|
220 |
-
|
221 |
-
|
222 |
-
|
223 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
224 |
|
225 |
demo.launch(share=True)
|
|
|
24 |
"filename" : "layer_31/width_16k/average_l0_76/params.npz"
|
25 |
}
|
26 |
|
27 |
+
title = """
|
28 |
+
<div class='parent' align="center">
|
29 |
+
<div class='child' style="display: inline-block !important; margin-bottom: 20px;">
|
30 |
+
<h1 style="margin-bottom: 30px;">🔍Interpretable Classifier for movie ratings using Gemma 2 with SAEs</h1>
|
31 |
+
</div>
|
32 |
+
</div>
|
33 |
+
<div class='parent' align="center">
|
34 |
+
<p>This space demonstrates how a linear classifier trained on top of features learned by Sparse Auto Encoders (SAEs) can be used to create interpretable natural language classifiers.</p>
|
35 |
+
<p>We leverage the interpretability API of <b>Neuronpedia</b> to provide more information about the features used by the LLM (like what tokens activate it the most and their distribution).</p>
|
36 |
+
<p><b>More resources on interpretability for LLMs using SAEs:</b></p>
|
37 |
+
</div>
|
38 |
+
<ul>
|
39 |
+
<li><a href="https://transformer-circuits.pub/2024/scaling-monosemanticity/">Anthropic: Scaling Monosemanticity: Extracting Interpretable Features from Claude 3 Sonnet</a></li>
|
40 |
+
<li><a href="https://blog.eleuther.ai/autointerp/">EleutherAI: Open Source Automated Interpretability for Sparse Autoencoder Features</a></li>
|
41 |
+
<li><a href="https://www.gemma.ai/gemma-scope">Gemma Scope: Open Sparse Autoencoders Everywhere All At Once on Gemma 2</a></li>
|
42 |
+
</ul>
|
43 |
+
<div class='parent' align="center">
|
44 |
+
<p>About us: <b> 🌊 LaVague</b> is an open-source framework to build AI Web Agents. Check out our <a href="https://github.com/lavague-ai/LaVague">GitHub</a> or join our <a href="https://discord.com/invite/SDxn9KpqX9">Discord</a>.</p>
|
45 |
+
</div>
|
46 |
+
"""
|
47 |
+
|
48 |
+
css = """
|
49 |
+
.my-button {
|
50 |
+
height: 100px; /* Increase the height of the buttons */
|
51 |
+
width: 100%; /* Make sure the button takes the full width */
|
52 |
+
max-width: 300px; /* Optional: set a max width */
|
53 |
+
max-height: 80px;
|
54 |
+
font-size: 1.1rem; /* Increase font size */
|
55 |
+
}
|
56 |
+
.button-container {
|
57 |
+
display: flex;
|
58 |
+
justify-content: center; /* Center buttons horizontally */
|
59 |
+
align-items: center; /* Center buttons vertically */
|
60 |
+
height: 100%; /* Ensure it takes up the full height */
|
61 |
+
width: 100%; /* Ensure it takes up the full width */
|
62 |
+
}
|
63 |
+
"""
|
64 |
+
|
65 |
C = 0.01
|
66 |
|
67 |
model_name = params["model_name"]
|
|
|
206 |
html = gr.HTML(html_content)
|
207 |
return html
|
208 |
|
209 |
+
with gr.Blocks(gr.themes.Default(primary_hue="blue", secondary_hue="neutral"), css=css) as demo:
|
210 |
+
with gr.Tab(""):
|
211 |
+
with gr.Row():
|
212 |
+
gr.HTML(title)
|
213 |
+
with gr.Row():
|
214 |
+
with gr.Column(scale=4):
|
215 |
+
input_text = gr.Textbox(label="Input", show_label=False, value=DEFAULT_EXAMPLE)
|
216 |
+
gr.Examples(
|
217 |
+
examples=examples,
|
218 |
+
inputs=input_text,
|
219 |
+
)
|
220 |
+
with gr.Column(scale=1):
|
221 |
+
run_button = gr.Button("Run")
|
222 |
+
|
223 |
+
with gr.Row():
|
224 |
+
label = gr.Label(label="Scores")
|
225 |
+
|
226 |
+
with gr.Row():
|
227 |
+
with gr.Column(scale=1):
|
228 |
+
plot = gr.Plot(label="Plot")
|
229 |
+
dropdown = gr.Dropdown(choices=["Option 1"], label="Features")
|
230 |
+
|
231 |
+
with gr.Column(scale=1):
|
232 |
+
highlighted_text = gr.HighlightedText(
|
233 |
+
label="Activating Tokens",
|
234 |
+
combine_adjacent=True,
|
235 |
+
show_legend=True,
|
236 |
+
color_map={"+": "red", "-": "green"})
|
237 |
+
|
238 |
+
with gr.Row():
|
239 |
+
html = gr.HTML()
|
240 |
+
|
241 |
+
# Connect the components
|
242 |
+
run_button.click(
|
243 |
+
fn=get_features,
|
244 |
+
inputs=[input_text],
|
245 |
+
outputs=[label, plot, dropdown],
|
246 |
+
).then(
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
247 |
fn=get_highlighted_text,
|
248 |
+
inputs=[input_text, dropdown],
|
249 |
+
outputs=[highlighted_text]
|
250 |
+
).then(
|
251 |
+
fn=get_feature_iframe,
|
252 |
+
inputs=[dropdown],
|
253 |
+
outputs=[html]
|
254 |
+
)
|
255 |
+
|
256 |
+
dropdown.change(
|
257 |
+
fn=get_highlighted_text,
|
258 |
+
inputs=[input_text, dropdown],
|
259 |
+
outputs=[highlighted_text]
|
260 |
+
).then(
|
261 |
+
fn=get_feature_iframe,
|
262 |
+
inputs=[dropdown],
|
263 |
+
outputs=[html]
|
264 |
+
)
|
265 |
|
266 |
demo.launch(share=True)
|