Update app.py
Browse files
app.py
CHANGED
@@ -40,7 +40,6 @@ suggested_interpretation_prompts = [
|
|
40 |
def initialize_gpu():
|
41 |
pass
|
42 |
|
43 |
-
|
44 |
def reset_model(model_name, *extra_components):
|
45 |
# extract model info
|
46 |
model_args = deepcopy(model_info[model_name])
|
@@ -133,7 +132,7 @@ with gr.Blocks(theme=gr.themes.Default(), css='styles.css') as demo:
|
|
133 |
'''
|
134 |
**👾 This space is a simple introduction to the emerging trend of models interpreting their OWN hidden states in free form natural language!!👾**
|
135 |
This idea was investigated in the paper **Patchscopes** ([Ghandeharioun et al., 2024](https://arxiv.org/abs/2401.06102)) and was further explored in **SelfIE** ([Chen et al., 2024](https://arxiv.org/abs/2403.10949)).
|
136 |
-
|
137 |
We will follow the SelfIE implementation in this space for concreteness. Patchscopes are so general that they encompass many other interpretation techniques too!!!
|
138 |
''', line_breaks=True)
|
139 |
|
@@ -204,10 +203,11 @@ with gr.Blocks(theme=gr.themes.Default(), css='styles.css') as demo:
|
|
204 |
model_chooser.change(reset_model, [model_chooser, *extra_components], extra_components)
|
205 |
|
206 |
for i, btn in enumerate(tokens_container):
|
207 |
-
btn.click(partial(run_interpretation, i=i),
|
208 |
-
|
209 |
-
|
210 |
-
|
|
|
211 |
|
212 |
original_prompt_btn.click(get_hidden_states,
|
213 |
[original_prompt_raw],
|
|
|
40 |
def initialize_gpu():
|
41 |
pass
|
42 |
|
|
|
43 |
def reset_model(model_name, *extra_components):
|
44 |
# extract model info
|
45 |
model_args = deepcopy(model_info[model_name])
|
|
|
132 |
'''
|
133 |
**👾 This space is a simple introduction to the emerging trend of models interpreting their OWN hidden states in free form natural language!!👾**
|
134 |
This idea was investigated in the paper **Patchscopes** ([Ghandeharioun et al., 2024](https://arxiv.org/abs/2401.06102)) and was further explored in **SelfIE** ([Chen et al., 2024](https://arxiv.org/abs/2403.10949)).
|
135 |
+
Honorary mention: **Speaking Probes** ([Dar, 2023](https://towardsdatascience.com/speaking-probes-self-interpreting-models-7a3dc6cb33d6) - my own work 🥳). It was less mature but had the same idea in mind. I think it's can be a great introduction to the subject!
|
136 |
We will follow the SelfIE implementation in this space for concreteness. Patchscopes are so general that they encompass many other interpretation techniques too!!!
|
137 |
''', line_breaks=True)
|
138 |
|
|
|
203 |
model_chooser.change(reset_model, [model_chooser, *extra_components], extra_components)
|
204 |
|
205 |
for i, btn in enumerate(tokens_container):
|
206 |
+
btn.click(partial(run_interpretation, i=i), dict(intepretation_prompt_raw=interpretation_prompt,
|
207 |
+
max_new_tokens=num_tokens, do_sample=do_sample, temperature=temperature,
|
208 |
+
top_k=top_k, top_p=top_p, repetition_penalty=repetition_penalty,
|
209 |
+
length_penalty=length_penalty,
|
210 |
+
), [progress_dummy, *interpretation_bubbles])
|
211 |
|
212 |
original_prompt_btn.click(get_hidden_states,
|
213 |
[original_prompt_raw],
|