Upload 2 files
Browse files- app.py +2 -2
- inflections_funcs.py +6 -7
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 |
|
|
|
|
| 63 |
@spaces.GPU
|
| 64 |
def predict(prompt, dark_mode, temperature):
|
| 65 |
"""
|
| 66 |
+
Generates responses for 5 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) % 5
|
| 99 |
main_html, detail_html = generate_beam_html(new_index, all_beams_data, dark_mode)
|
| 100 |
return main_html, detail_html, new_index
|
| 101 |
|
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."},
|
|
@@ -38,15 +38,14 @@ 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 |
-
diversity_penalty=1.0, # Penalty to encourage diversity between groups
|
| 44 |
-
num_return_sequences=3, # Returns the top 3 results
|
| 45 |
return_dict_in_generate=True,
|
| 46 |
output_scores=True,
|
| 47 |
temperature=temperature,
|
| 48 |
-
|
| 49 |
-
|
|
|
|
| 50 |
|
| 51 |
transcription = processor.batch_decode(generated_dicts.sequences, skip_special_tokens=True)
|
| 52 |
|
|
|
|
| 19 |
|
| 20 |
def make_beams(model: AutoModelForCausalLM, processor: AutoProcessor, initial_prompt: str, temperature: float = 1.0) -> Tuple[Any, List[str]]:
|
| 21 |
'''
|
| 22 |
+
Generates 5 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=5, # Use Beam Sampling for better diversity
|
| 42 |
+
num_return_sequences=5, # Returns 5 distinct candidates
|
|
|
|
|
|
|
| 43 |
return_dict_in_generate=True,
|
| 44 |
output_scores=True,
|
| 45 |
temperature=temperature,
|
| 46 |
+
top_p=0.9, # Nucleus sampling to increase diversity
|
| 47 |
+
top_k=50, # Top-K sampling to filter noise
|
| 48 |
+
do_sample=True if temperature > 0 else False)
|
| 49 |
|
| 50 |
transcription = processor.batch_decode(generated_dicts.sequences, skip_special_tokens=True)
|
| 51 |
|