import evaluate import gradio as gr import os # Create your own Gradio interface instead of using the built-in widget def create_interface(module): def evaluate_fn(prediction, references, pos_pred, neg_pred): # Check if all required fields are filled if not prediction or prediction.strip() == "": return "", "", "", "Please provide a candidate hypothesis to evaluate." if not references or references.strip() == "": return "", "", "", "Please provide a validation program." if not pos_pred or pos_pred.strip() == "": return "", "", "", "Please specify the positive predicate name." if not neg_pred or neg_pred.strip() == "": return "", "", "", "Please specify the negative predicate name." # Process a single prediction instead of multiple pred = prediction.strip() # Create reference with evaluation_config at the top level ref = { "validation_program": references.strip(), "evaluation_config": { "positive_predicate": pos_pred, "negative_predicate": neg_pred } } # Use a list with a single prediction and reference results = module.compute(predictions=[pred], references=[ref]) # Extract the error message from detailed_results if it exists error_msg = "" if results["detailed_results"] and len(results["detailed_results"]) > 0: error = results["detailed_results"][0].get("error") if error: error_msg = error return ( f"Accuracy score: {results['accuracy']:.4f}", f"Partial score: {results['partial_score']:.4f}", f"Syntax score: {results['syntax_score']:.4f}", error_msg ) # Helper function to load example data def load_example(example): return ( example["rule"], example["validation"], example["pos_pred"], example["neg_pred"] ) # Read README.md content readme_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "README.md") with open(readme_path, 'r') as f: readme_content = f.read() readme_content = '# Metric Card ' + readme_content.split('# Metric Card ')[1] # Define examples for quick use example_train = { "description": "Basic Train Problem", "validation": """eastbound(train0). has_car(train0, car0_1). car_num(car0_1, 1). car_color(car0_1, white). car_len(car0_1, short). has_wall(car0_1, full). westbound(train1). has_car(train1, car1_1). car_num(car1_1, 1). car_color(car1_1, yellow). car_len(car1_1, short). has_wall(car1_1, full). """, "rule": "eastbound(Train):- has_car(Train, Car1), car_color(Car1, white).", "pos_pred": "eastbound", "neg_pred": "westbound" } example_family = { "description": "Family Relationships", "validation": """% Custom problem parent(john, mary). parent(john, bob). parent(alice, bob). parent(susan, alice). % Examples grandparent(susan, bob). not_grandparent(john, alice).""", "rule": "grandparent(X, Y) :- parent(X, Z), parent(Z, Y).", "pos_pred": "grandparent", "neg_pred": "not_grandparent" } with gr.Blocks(title="Symbolic Judge") as demo: with gr.Tab("Evaluation"): gr.Markdown("# Symbolic Judge: Verifiable Rewards for Scalable Logical Reasoning") gr.Markdown(""" Verifiable Rewards for Scalable Logical Reasoning (**SLR**) introduces a **symbolic judge** that provides verifiable rewards for logical reasoning tasks. To check whether a given task is solved, the symbolic judge evaluates a candidate solution (i.e., a logic rule, typically generated by a language model) and using an **executable validation program** that encodes the task's background knowledge and labeled examples. Evaluations performed by the symbolic judge are fully verifiable and grounded in formal logic, ensuring an automatic, transparent, and reproducible standard for evaluation and reward in both supervised and reinforcement learning settings. ### How it Works - **Input:** The symbolic judge takes as input a candidate hypothesis (logic rule) and an executable validation program containing background knowledge and examples. - **Execution:** The candidate rule is executed against the validation program using a Prolog interpreter. - **Correctness Criteria:** The rule is considered correct if it entails all positive examples and rejects all negative examples. - **Metrics:** The symbolic judge computes a range of evaluation metrics (detailed below). - **Usage:** see **Documentation tab** for details on how to use the symbolic judge in your own projects. **Note:** A local Prolog interpreter is required to execute validation programs. """) with gr.Row(): with gr.Column(scale=1): with gr.Group(): gr.Markdown("### Model Output") prediction_input = gr.Textbox( label="Candidate Hypothesis to be evaluated(predicted rule by the model)", placeholder="eastbound(T) :- has_car(T, C), short(C), open(C).", lines=5 ) with gr.Group(): gr.Markdown("### Validation Program") references_input = gr.Textbox( label="The validation program contains background knowledge and examples for testing", placeholder="% Background knowledge\ncar(car_1). car(car_2).\nlong(car_2). short(car_1).\nopen(car_1). closed(car_2).\n\n% Examples\neastbound(train_1).\nwestbound(train_2).\n\n% Train configurations\nhas_car(train_1, car_1).\nhas_car(train_2, car_2).", lines=12 ) with gr.Row(): pos_pred_input = gr.Textbox( label="Positive Validation Examples", value="eastbound", placeholder="eastbound", info="The predicate name identifying positive examples in the validation program" ) neg_pred_input = gr.Textbox( label="Negative Validation Examples", value="westbound", placeholder="westbound", info="The predicate name identifying negative examples in the validation program" ) eval_button = gr.Button("Evaluate", variant="primary") with gr.Column(scale=1): with gr.Group(): gr.Markdown("### Evaluation Metrics") with gr.Group(): accuracy_output = gr.Textbox( label="Overall Accuracy", info="Proportion of hypotheses that solve the tasks", container=True ) partial_score_output = gr.Textbox( label="Partial Score", info="Proportion of examples that are correctly classified in the tasks", container=True ) syntax_score_output = gr.Textbox( label="Syntax Score", info="Proportion of syntactically valid hypothesis", container=True ) error_output = gr.Textbox( label="Syntax Details", info="Error messages for syntax errors or execution failures", container=True, ) gr.Markdown("Note: This interface evaluates a single hypothesis at a time. Use Python API for batch processing") # Add the examples section # Example list examples = [ ["Train Problem", example_train], ["Family Relationships", example_family] ] with gr.Accordion("Example Logical Reasoning Tasks", open=True): example_radio = gr.Radio([ex[0] for ex in examples], label="Select an example", value="Train Problem") # Show preview of selected example with gr.Row(): with gr.Column(): gr.Markdown("### Selected Example Preview") example_description = gr.Markdown("**Description**: " + example_train["description"]) with gr.Row(): with gr.Column(): gr.Markdown("#### Candidate Hypothesis") example_rule = gr.Code(value=example_train["rule"]) with gr.Column(): gr.Markdown("#### Validation Program") example_validation = gr.Code(value=example_train["validation"]) with gr.Row(): with gr.Column(): gr.Markdown("#### Validation Examples") example_predicates = gr.Markdown(f""" **Positive Examples**: `{example_train["pos_pred"]}` **Negative Examples**: `{example_train["neg_pred"]}` """) # Load button for the selected example load_button = gr.Button("Load Selected Example", variant="secondary") gr.Markdown("### Citation") gr.Markdown(""" If you use Symbolic Judge in your work, please cite: ``` @misc{helff2025slrautomatedsynthesisframework, title={SLR: An Automated Synthesis Framework for Scalable Logical Reasoning}, author={Lukas Helff and Ahmad Omar and Felix Friedrich and Wolfgang Stammer and Antonia Wüst and Tim Woydt and Rupert Mitchell and Patrick Schramowski and Kristian Kersting}, year={2025}, eprint={2506.15787}, archivePrefix={arXiv}, primaryClass={cs.AI}, url={https://arxiv.org/abs/2506.15787}, } ``` """) # Set up event handlers for the example selection def update_example_preview(selection): selected_example = next((ex[1] for ex in examples if ex[0] == selection), example_train) return ( "**Description**: " + selected_example["description"], selected_example["rule"], selected_example["validation"], f""" **Positive Examples**: `{selected_example["pos_pred"]}` **Negative Examples**: `{selected_example["neg_pred"]}` """ ) example_radio.change( fn=update_example_preview, inputs=[example_radio], outputs=[example_description, example_rule, example_validation, example_predicates] ) # Event handler for the load button def load_selected_example(selection): selected_example = next((ex[1] for ex in examples if ex[0] == selection), example_train) return load_example(selected_example) load_button.click( fn=load_selected_example, inputs=[example_radio], outputs=[prediction_input, references_input, pos_pred_input, neg_pred_input] ) # Set up the evaluate button eval_button.click( fn=evaluate_fn, inputs=[prediction_input, references_input, pos_pred_input, neg_pred_input], outputs=[accuracy_output, partial_score_output, syntax_score_output, error_output] ) with gr.Tab("Documentation"): gr.Markdown(readme_content) return demo # Use a local path instead of a module name module = evaluate.load("AIML-TUDA/VerifiableRewardsForScalableLogicalReasoning") create_interface(module).launch()