Upload 2 files
Browse files- app.py +3 -2
- inflections_funcs.py +7 -4
app.py
CHANGED
|
@@ -63,7 +63,7 @@ def generate_beam_html(index, all_beams_data, dark_mode):
|
|
| 63 |
@spaces.GPU
|
| 64 |
def predict(prompt, dark_mode, temperature):
|
| 65 |
"""
|
| 66 |
-
Generates responses for
|
| 67 |
"""
|
| 68 |
# Generate beams
|
| 69 |
generated_dicts, transcription = make_beams(model, processor, prompt, temperature=temperature)
|
|
@@ -95,7 +95,7 @@ def switch_beam(current_index, all_beams_data, dark_mode):
|
|
| 95 |
if all_beams_data is None:
|
| 96 |
return None, None, 0
|
| 97 |
|
| 98 |
-
new_index = (current_index + 1) %
|
| 99 |
main_html, detail_html = generate_beam_html(new_index, all_beams_data, dark_mode)
|
| 100 |
return main_html, detail_html, new_index
|
| 101 |
|
|
@@ -111,6 +111,7 @@ with gr.Blocks() as demo:
|
|
| 111 |
- Display the first output beam, highlighting any word or parts of words (tokens) with a probability score < 0.6 in red.
|
| 112 |
- These unconfident words/tokens can be considered inflection points, or places where the output could change easily.
|
| 113 |
- Provide a toggle button to view detailed word/token probabilities for the beam.
|
|
|
|
| 114 |
''')
|
| 115 |
|
| 116 |
with gr.Column():
|
|
|
|
| 63 |
@spaces.GPU
|
| 64 |
def predict(prompt, dark_mode, temperature):
|
| 65 |
"""
|
| 66 |
+
Generates responses for 6 beams and returns the first beam's visualization and visibility for controls.
|
| 67 |
"""
|
| 68 |
# Generate beams
|
| 69 |
generated_dicts, transcription = make_beams(model, processor, prompt, temperature=temperature)
|
|
|
|
| 95 |
if all_beams_data is None:
|
| 96 |
return None, None, 0
|
| 97 |
|
| 98 |
+
new_index = (current_index + 1) % 6
|
| 99 |
main_html, detail_html = generate_beam_html(new_index, all_beams_data, dark_mode)
|
| 100 |
return main_html, detail_html, new_index
|
| 101 |
|
|
|
|
| 111 |
- Display the first output beam, highlighting any word or parts of words (tokens) with a probability score < 0.6 in red.
|
| 112 |
- These unconfident words/tokens can be considered inflection points, or places where the output could change easily.
|
| 113 |
- Provide a toggle button to view detailed word/token probabilities for the beam.
|
| 114 |
+
- Switch between beams to see different possible outputs and their associated token probabilities.
|
| 115 |
''')
|
| 116 |
|
| 117 |
with gr.Column():
|
inflections_funcs.py
CHANGED
|
@@ -19,7 +19,7 @@ def start_model(model_id: str = "google/gemma-4-31B-it"):
|
|
| 19 |
|
| 20 |
def make_beams(model: AutoModelForCausalLM, processor: AutoProcessor, initial_prompt: str, temperature: float = 1.0) -> Tuple[Any, List[str]]:
|
| 21 |
'''
|
| 22 |
-
Generates
|
| 23 |
'''
|
| 24 |
messages = [
|
| 25 |
{"role": "system", "content": "You are a helpful assistant."},
|
|
@@ -38,12 +38,15 @@ def make_beams(model: AutoModelForCausalLM, processor: AutoProcessor, initial_pr
|
|
| 38 |
# Generate output
|
| 39 |
generated_dicts = model.generate(**inputs,
|
| 40 |
max_new_tokens=1024,
|
| 41 |
-
num_beams=
|
| 42 |
-
|
|
|
|
|
|
|
| 43 |
return_dict_in_generate=True,
|
| 44 |
output_scores=True,
|
| 45 |
temperature=temperature,
|
| 46 |
-
do_sample=True if temperature > 0 else False
|
|
|
|
| 47 |
|
| 48 |
transcription = processor.batch_decode(generated_dicts.sequences, skip_special_tokens=True)
|
| 49 |
|
|
|
|
| 19 |
|
| 20 |
def make_beams(model: AutoModelForCausalLM, processor: AutoProcessor, initial_prompt: str, temperature: float = 1.0) -> Tuple[Any, List[str]]:
|
| 21 |
'''
|
| 22 |
+
Generates 6 diverse beams in response to a prompt.
|
| 23 |
'''
|
| 24 |
messages = [
|
| 25 |
{"role": "system", "content": "You are a helpful assistant."},
|
|
|
|
| 38 |
# Generate output
|
| 39 |
generated_dicts = model.generate(**inputs,
|
| 40 |
max_new_tokens=1024,
|
| 41 |
+
num_beams=6, # Enables diverse beam search with 6 active beams
|
| 42 |
+
num_beam_groups=3, # Divide beams into 3 groups for diversity
|
| 43 |
+
diversity_penalty=1.0, # Penalty to encourage diversity between groups
|
| 44 |
+
num_return_sequences=6, # Returns the top 6 results
|
| 45 |
return_dict_in_generate=True,
|
| 46 |
output_scores=True,
|
| 47 |
temperature=temperature,
|
| 48 |
+
do_sample=True if temperature > 0 else False,
|
| 49 |
+
trust_remote_code=True)
|
| 50 |
|
| 51 |
transcription = processor.batch_decode(generated_dicts.sequences, skip_special_tokens=True)
|
| 52 |
|