Upload app.py with huggingface_hub
Browse files
app.py
ADDED
|
@@ -0,0 +1,168 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
import os
|
| 2 |
+
|
| 3 |
+
import evaluate
|
| 4 |
+
import gradio as gr
|
| 5 |
+
|
| 6 |
+
|
| 7 |
+
def create_interface(module):
|
| 8 |
+
def evaluate_fn(prediction, validation_program, pos_pred, neg_pred):
|
| 9 |
+
if not prediction or not prediction.strip():
|
| 10 |
+
return "", "", "", "", "", "Please provide a candidate hypothesis."
|
| 11 |
+
if not validation_program or not validation_program.strip():
|
| 12 |
+
return "", "", "", "", "", "Please provide a validation program."
|
| 13 |
+
if not pos_pred or not pos_pred.strip():
|
| 14 |
+
return "", "", "", "", "", "Please specify the positive predicate."
|
| 15 |
+
if not neg_pred or not neg_pred.strip():
|
| 16 |
+
return "", "", "", "", "", "Please specify the negative predicate."
|
| 17 |
+
|
| 18 |
+
ref = {
|
| 19 |
+
"validation_program": validation_program.strip(),
|
| 20 |
+
"evaluation_config": {
|
| 21 |
+
"positive_predicate": pos_pred.strip(),
|
| 22 |
+
"negative_predicate": neg_pred.strip(),
|
| 23 |
+
},
|
| 24 |
+
}
|
| 25 |
+
results = module.compute(
|
| 26 |
+
predictions=[prediction.strip()],
|
| 27 |
+
references=[ref],
|
| 28 |
+
verbose=False,
|
| 29 |
+
)
|
| 30 |
+
|
| 31 |
+
d = results["detailed_results"][0]
|
| 32 |
+
error_msg = d.get("error") or ""
|
| 33 |
+
|
| 34 |
+
ext_icon = "β
" if d["extensional_correct"] else "β"
|
| 35 |
+
iso_icon = "β
" if d["isomorphic_correct"] else "β"
|
| 36 |
+
shortcut_icon = "β οΈ Reward shortcut detected" if d["is_reward_shortcut"] else "β No shortcut"
|
| 37 |
+
|
| 38 |
+
return (
|
| 39 |
+
f"{ext_icon} {results['extensional_accuracy']:.4f} (partial: {d['extensional_partial']:.4f})",
|
| 40 |
+
f"{iso_icon} {results['isomorphic_accuracy']:.4f} (partial: {d['isomorphic_partial']:.4f})",
|
| 41 |
+
shortcut_icon,
|
| 42 |
+
f"{results['syntax_score']:.4f}",
|
| 43 |
+
error_msg,
|
| 44 |
+
)
|
| 45 |
+
|
| 46 |
+
# ------------------------------------------------------------------ #
|
| 47 |
+
# Examples
|
| 48 |
+
# ------------------------------------------------------------------ #
|
| 49 |
+
EXAMPLES = {
|
| 50 |
+
"Genuine rule": {
|
| 51 |
+
"description": "A genuine relational rule β passes both verifications.",
|
| 52 |
+
"rule": "eastbound(Train) :- has_car(Train, Car), car_color(Car, red).",
|
| 53 |
+
"validation": (
|
| 54 |
+
"eastbound(train0).\nhas_car(train0, car0_1).\ncar_color(car0_1, red).\n\n"
|
| 55 |
+
"westbound(train1).\nhas_car(train1, car1_1).\ncar_color(car1_1, blue).\n\n"
|
| 56 |
+
"eastbound(train2).\nhas_car(train2, car2_1).\ncar_color(car2_1, red).\n\n"
|
| 57 |
+
"westbound(train3).\nhas_car(train3, car3_1).\ncar_color(car3_1, blue).\n"
|
| 58 |
+
),
|
| 59 |
+
"pos_pred": "eastbound",
|
| 60 |
+
"neg_pred": "westbound",
|
| 61 |
+
},
|
| 62 |
+
"Blatant shortcut": {
|
| 63 |
+
"description": "Grounded enumeration β passes extensional, fails isomorphic.",
|
| 64 |
+
"rule": "eastbound(train0). eastbound(train2).",
|
| 65 |
+
"validation": (
|
| 66 |
+
"eastbound(train0).\nhas_car(train0, car0_1).\ncar_color(car0_1, red).\n\n"
|
| 67 |
+
"westbound(train1).\nhas_car(train1, car1_1).\ncar_color(car1_1, blue).\n\n"
|
| 68 |
+
"eastbound(train2).\nhas_car(train2, car2_1).\ncar_color(car2_1, red).\n\n"
|
| 69 |
+
"westbound(train3).\nhas_car(train3, car3_1).\ncar_color(car3_1, blue).\n"
|
| 70 |
+
),
|
| 71 |
+
"pos_pred": "eastbound",
|
| 72 |
+
"neg_pred": "westbound",
|
| 73 |
+
},
|
| 74 |
+
"Negation shortcut": {
|
| 75 |
+
"description": "Uses \\+ westbound β passes extensional via bridge rule, fails isomorphic.",
|
| 76 |
+
"rule": "eastbound(T) :- \\+ westbound(T).",
|
| 77 |
+
"validation": (
|
| 78 |
+
"eastbound(train0).\nhas_car(train0, car0_1).\ncar_color(car0_1, red).\n\n"
|
| 79 |
+
"westbound(train1).\nhas_car(train1, car1_1).\ncar_color(car1_1, blue).\n\n"
|
| 80 |
+
"eastbound(train2).\nhas_car(train2, car2_1).\ncar_color(car2_1, red).\n\n"
|
| 81 |
+
"westbound(train3).\nhas_car(train3, car3_1).\ncar_color(car3_1, blue).\n"
|
| 82 |
+
),
|
| 83 |
+
"pos_pred": "eastbound",
|
| 84 |
+
"neg_pred": "westbound",
|
| 85 |
+
},
|
| 86 |
+
}
|
| 87 |
+
|
| 88 |
+
readme_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "README.md")
|
| 89 |
+
with open(readme_path) as f:
|
| 90 |
+
readme = f.read()
|
| 91 |
+
|
| 92 |
+
def update_preview(name):
|
| 93 |
+
ex = EXAMPLES[name]
|
| 94 |
+
return (
|
| 95 |
+
f"**{ex['description']}**",
|
| 96 |
+
ex["rule"],
|
| 97 |
+
ex["validation"],
|
| 98 |
+
f"`{ex['pos_pred']}` / `{ex['neg_pred']}`",
|
| 99 |
+
)
|
| 100 |
+
|
| 101 |
+
def load_example(name):
|
| 102 |
+
ex = EXAMPLES[name]
|
| 103 |
+
return ex["rule"], ex["validation"], ex["pos_pred"], ex["neg_pred"]
|
| 104 |
+
|
| 105 |
+
with gr.Blocks(title="Isomorphic Perturbation Testing") as demo:
|
| 106 |
+
with gr.Tab("Evaluate"):
|
| 107 |
+
gr.Markdown("# Isomorphic Perturbation Testing (IPT)")
|
| 108 |
+
gr.Markdown(
|
| 109 |
+
"Diagnose whether a model output is a **genuine rule** or a **reward shortcut** "
|
| 110 |
+
"by running both extensional and isomorphic verification. "
|
| 111 |
+
"A shortcut passes extensional (original object names) but fails isomorphic "
|
| 112 |
+
"(object constants bijectively renamed)."
|
| 113 |
+
)
|
| 114 |
+
|
| 115 |
+
with gr.Row():
|
| 116 |
+
with gr.Column():
|
| 117 |
+
prediction_input = gr.Textbox(
|
| 118 |
+
label="Candidate Hypothesis (model output)",
|
| 119 |
+
placeholder="eastbound(T) :- has_car(T, C), car_color(C, red).",
|
| 120 |
+
lines=4,
|
| 121 |
+
)
|
| 122 |
+
validation_input = gr.Textbox(
|
| 123 |
+
label="Validation Program",
|
| 124 |
+
placeholder="eastbound(train0).\nhas_car(train0, car0_1).\n...",
|
| 125 |
+
lines=10,
|
| 126 |
+
)
|
| 127 |
+
with gr.Row():
|
| 128 |
+
pos_pred_input = gr.Textbox(label="Positive predicate", value="eastbound")
|
| 129 |
+
neg_pred_input = gr.Textbox(label="Negative predicate", value="westbound")
|
| 130 |
+
eval_btn = gr.Button("Evaluate", variant="primary")
|
| 131 |
+
|
| 132 |
+
with gr.Column():
|
| 133 |
+
gr.Markdown("### Results")
|
| 134 |
+
ext_out = gr.Textbox(label="Extensional verification")
|
| 135 |
+
iso_out = gr.Textbox(label="Isomorphic verification")
|
| 136 |
+
shortcut_out = gr.Textbox(label="Shortcut verdict")
|
| 137 |
+
syntax_out = gr.Textbox(label="Syntax score")
|
| 138 |
+
error_out = gr.Textbox(label="Error / warnings")
|
| 139 |
+
gr.Markdown(
|
| 140 |
+
"_This interface evaluates one hypothesis at a time. "
|
| 141 |
+
"Use the Python API for batch processing._"
|
| 142 |
+
)
|
| 143 |
+
|
| 144 |
+
with gr.Accordion("Examples", open=True):
|
| 145 |
+
example_radio = gr.Radio(list(EXAMPLES), label="Select example", value="Genuine rule")
|
| 146 |
+
example_desc = gr.Markdown(f"**{EXAMPLES['Genuine rule']['description']}**")
|
| 147 |
+
with gr.Row():
|
| 148 |
+
example_rule_view = gr.Code(value=EXAMPLES["Genuine rule"]["rule"], label="Rule")
|
| 149 |
+
example_vp_view = gr.Code(value=EXAMPLES["Genuine rule"]["validation"], label="Validation program")
|
| 150 |
+
example_preds = gr.Markdown(f"`eastbound` / `westbound`")
|
| 151 |
+
load_btn = gr.Button("Load example", variant="secondary")
|
| 152 |
+
|
| 153 |
+
example_radio.change(update_preview, example_radio,
|
| 154 |
+
[example_desc, example_rule_view, example_vp_view, example_preds])
|
| 155 |
+
load_btn.click(load_example, example_radio,
|
| 156 |
+
[prediction_input, validation_input, pos_pred_input, neg_pred_input])
|
| 157 |
+
eval_btn.click(evaluate_fn,
|
| 158 |
+
[prediction_input, validation_input, pos_pred_input, neg_pred_input],
|
| 159 |
+
[ext_out, iso_out, shortcut_out, syntax_out, error_out])
|
| 160 |
+
|
| 161 |
+
with gr.Tab("Documentation"):
|
| 162 |
+
gr.Markdown(readme)
|
| 163 |
+
|
| 164 |
+
return demo
|
| 165 |
+
|
| 166 |
+
|
| 167 |
+
module = evaluate.load("AIML-TUDA/IsomorphicPerturbationTesting")
|
| 168 |
+
create_interface(module).launch()
|