Spaces:
Sleeping
Sleeping
Create app_models_config.py
Browse files- app_models_config.py +167 -0
app_models_config.py
ADDED
|
@@ -0,0 +1,167 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import gradio as gr
|
| 2 |
+
|
| 3 |
+
# Available models
|
| 4 |
+
AVAILABLE_MODELS = [
|
| 5 |
+
"(Select Model)",
|
| 6 |
+
"mistralai/Mistral-7B-v0.1",
|
| 7 |
+
"meta-llama/Llama-2-7b-hf",
|
| 8 |
+
"google/flan-t5-xxl",
|
| 9 |
+
"tiiuae/falcon-7b"
|
| 10 |
+
]
|
| 11 |
+
|
| 12 |
+
def create_model_config_section():
|
| 13 |
+
"""
|
| 14 |
+
Creates the "Head to Head - Choose Models" section with two model configurations side by side.
|
| 15 |
+
Returns the components needed for the main app.
|
| 16 |
+
"""
|
| 17 |
+
with gr.Column() as model_config_container:
|
| 18 |
+
gr.Markdown("## (C) Head to Head - Choose Models to evaluate against each other")
|
| 19 |
+
|
| 20 |
+
with gr.Row():
|
| 21 |
+
# Left column - Model 1 configuration
|
| 22 |
+
with gr.Column(scale=1) as model1_column:
|
| 23 |
+
with gr.Group(elem_classes=["config-box"]):
|
| 24 |
+
gr.Markdown("### Model 1")
|
| 25 |
+
|
| 26 |
+
model1_dropdown = gr.Dropdown(
|
| 27 |
+
choices=AVAILABLE_MODELS,
|
| 28 |
+
value="(Select Model)",
|
| 29 |
+
label="Select Model 1",
|
| 30 |
+
info="Choose the first model for head-to-head comparison"
|
| 31 |
+
)
|
| 32 |
+
|
| 33 |
+
model1_shots = gr.Slider(
|
| 34 |
+
minimum=0,
|
| 35 |
+
maximum=5,
|
| 36 |
+
value=5,
|
| 37 |
+
step=1,
|
| 38 |
+
label="Number of Few-shot Examples",
|
| 39 |
+
info="Number of examples to use for few-shot learning (0-5)"
|
| 40 |
+
)
|
| 41 |
+
|
| 42 |
+
model1_regex = gr.Textbox(
|
| 43 |
+
label="Regex Pattern",
|
| 44 |
+
placeholder="Optional: Apply regex pattern to model outputs",
|
| 45 |
+
info="Leave empty for no regex pattern"
|
| 46 |
+
)
|
| 47 |
+
|
| 48 |
+
model1_flash_attn = gr.Checkbox(
|
| 49 |
+
label="Use FlashAttention",
|
| 50 |
+
value=True,
|
| 51 |
+
info="Use FlashAttention for better performance (if supported by model)"
|
| 52 |
+
)
|
| 53 |
+
|
| 54 |
+
# Divider in the middle
|
| 55 |
+
with gr.Column(scale=0.1):
|
| 56 |
+
gr.Markdown('<div style="border-left: 1px solid #ddd; height: 100%;"></div>', elem_classes=["center-divider"])
|
| 57 |
+
|
| 58 |
+
# Right column - Model 2 configuration
|
| 59 |
+
with gr.Column(scale=1) as model2_column:
|
| 60 |
+
with gr.Group(elem_classes=["config-box"]):
|
| 61 |
+
gr.Markdown("### Model 2")
|
| 62 |
+
|
| 63 |
+
model2_dropdown = gr.Dropdown(
|
| 64 |
+
choices=AVAILABLE_MODELS,
|
| 65 |
+
value="(Select Model)",
|
| 66 |
+
label="Select Model 2",
|
| 67 |
+
info="Choose the second model for head-to-head comparison"
|
| 68 |
+
)
|
| 69 |
+
|
| 70 |
+
model2_shots = gr.Slider(
|
| 71 |
+
minimum=0,
|
| 72 |
+
maximum=5,
|
| 73 |
+
value=5,
|
| 74 |
+
step=1,
|
| 75 |
+
label="Number of Few-shot Examples",
|
| 76 |
+
info="Number of examples to use for few-shot learning (0-5)"
|
| 77 |
+
)
|
| 78 |
+
|
| 79 |
+
model2_regex = gr.Textbox(
|
| 80 |
+
label="Regex Pattern",
|
| 81 |
+
placeholder="Optional: Apply regex pattern to model outputs",
|
| 82 |
+
info="Leave empty for no regex pattern"
|
| 83 |
+
)
|
| 84 |
+
|
| 85 |
+
model2_flash_attn = gr.Checkbox(
|
| 86 |
+
label="Use FlashAttention",
|
| 87 |
+
value=True,
|
| 88 |
+
info="Use FlashAttention for better performance (if supported by model)"
|
| 89 |
+
)
|
| 90 |
+
|
| 91 |
+
# Error message area - initially hidden
|
| 92 |
+
model_config_error = gr.Markdown(
|
| 93 |
+
visible=False,
|
| 94 |
+
value="⚠️ **Error**: Both models and configurations are identical. Please select different models or configurations for comparison.",
|
| 95 |
+
elem_classes=["error-message"]
|
| 96 |
+
)
|
| 97 |
+
|
| 98 |
+
return {
|
| 99 |
+
'container': model_config_container,
|
| 100 |
+
'model1_dropdown': model1_dropdown,
|
| 101 |
+
'model1_shots': model1_shots,
|
| 102 |
+
'model1_regex': model1_regex,
|
| 103 |
+
'model1_flash_attn': model1_flash_attn,
|
| 104 |
+
'model2_dropdown': model2_dropdown,
|
| 105 |
+
'model2_shots': model2_shots,
|
| 106 |
+
'model2_regex': model2_regex,
|
| 107 |
+
'model2_flash_attn': model2_flash_attn,
|
| 108 |
+
'error_message': model_config_error
|
| 109 |
+
}
|
| 110 |
+
|
| 111 |
+
def validate_model_configs(model1, model1_shots, model1_regex, model1_flash,
|
| 112 |
+
model2, model2_shots, model2_regex, model2_flash):
|
| 113 |
+
"""
|
| 114 |
+
Validates that the two model configurations are not identical.
|
| 115 |
+
Returns:
|
| 116 |
+
- bool: Whether the configurations are valid (not identical)
|
| 117 |
+
- str: Error message if invalid, otherwise empty string
|
| 118 |
+
"""
|
| 119 |
+
if model1 == "(Select Model)" or model2 == "(Select Model)":
|
| 120 |
+
return True, ""
|
| 121 |
+
|
| 122 |
+
# Check if models and all configs are identical
|
| 123 |
+
if (model1 == model2 and
|
| 124 |
+
model1_shots == model2_shots and
|
| 125 |
+
model1_regex == model2_regex and
|
| 126 |
+
model1_flash == model2_flash):
|
| 127 |
+
return False, "⚠️ **Error**: Both models and configurations are identical. Please select different models or configurations for comparison."
|
| 128 |
+
|
| 129 |
+
return True, ""
|
| 130 |
+
|
| 131 |
+
def update_eval_button_state(model1, model1_shots, model1_regex, model1_flash,
|
| 132 |
+
model2, model2_shots, model2_regex, model2_flash):
|
| 133 |
+
"""
|
| 134 |
+
Checks model configurations and updates the error message visibility and eval button state.
|
| 135 |
+
"""
|
| 136 |
+
is_valid, error_msg = validate_model_configs(
|
| 137 |
+
model1, model1_shots, model1_regex, model1_flash,
|
| 138 |
+
model2, model2_shots, model2_regex, model2_flash
|
| 139 |
+
)
|
| 140 |
+
|
| 141 |
+
if model1 == "(Select Model)" or model2 == "(Select Model)":
|
| 142 |
+
return gr.update(visible=False), gr.update(interactive=False)
|
| 143 |
+
|
| 144 |
+
if not is_valid:
|
| 145 |
+
return gr.update(visible=True, value=error_msg), gr.update(interactive=False)
|
| 146 |
+
|
| 147 |
+
return gr.update(visible=False), gr.update(interactive=True)
|
| 148 |
+
|
| 149 |
+
def get_model_configs(model1, model1_shots, model1_regex, model1_flash,
|
| 150 |
+
model2, model2_shots, model2_regex, model2_flash):
|
| 151 |
+
"""
|
| 152 |
+
Returns the model configurations as structured data for the evaluation function.
|
| 153 |
+
"""
|
| 154 |
+
return {
|
| 155 |
+
"model1": {
|
| 156 |
+
"name": model1,
|
| 157 |
+
"shots": model1_shots,
|
| 158 |
+
"regex": model1_regex,
|
| 159 |
+
"flash_attention": model1_flash
|
| 160 |
+
},
|
| 161 |
+
"model2": {
|
| 162 |
+
"name": model2,
|
| 163 |
+
"shots": model2_shots,
|
| 164 |
+
"regex": model2_regex,
|
| 165 |
+
"flash_attention": model2_flash
|
| 166 |
+
}
|
| 167 |
+
}
|