File size: 13,031 Bytes
193db9d
 
 
 
 
 
 
 
 
 
973519b
193db9d
 
 
3b39b49
193db9d
3b39b49
 
193db9d
3b39b49
193db9d
3b39b49
193db9d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b39b49
193db9d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b39b49
 
 
193db9d
3b39b49
193db9d
 
 
 
 
 
 
 
 
 
 
 
3b39b49
193db9d
 
 
 
 
 
 
3b39b49
193db9d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b39b49
 
 
 
 
 
 
193db9d
3b39b49
 
 
 
 
 
193db9d
 
 
 
 
973519b
 
 
 
 
193db9d
 
 
 
 
 
 
3b39b49
193db9d
 
 
 
 
3b39b49
193db9d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b39b49
193db9d
3b39b49
193db9d
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
3b39b49
193db9d
3b39b49
193db9d
 
 
 
 
3b39b49
193db9d
 
 
 
 
 
 
 
 
 
 
 
 
3b39b49
193db9d
 
 
 
 
973519b
193db9d
 
 
 
 
 
 
 
3b39b49
193db9d
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
import json
import logging
from typing import Any

import gradio as gr
import pandas as pd
from datasets import Dataset

from components.model_pipeline.model_pipeline import PipelineInterface, PipelineState
from submission import submit
from workflows.qb.multi_step_agent import MultiStepBonusAgent
from workflows.qb.simple_agent import SimpleBonusAgent
from workflows.structs import ModelStep, Workflow

from .commons import get_qid_selector
from .plotting import (
    create_bonus_confidence_plot,
    create_bonus_html,
    create_scatter_pyplot,
    update_tossup_plot,
)
from .utils import evaluate_prediction


def process_bonus_results(results: list[dict]) -> pd.DataFrame:
    """Process results from bonus mode and prepare visualization data."""
    return pd.DataFrame(
        [
            {
                "Part": f"Part {r['part_number']}",
                "Correct?": "✅" if r["score"] == 1 else "❌",
                "Confidence": r["confidence"],
                "Prediction": r["answer"],
                "Explanation": r["explanation"],
            }
            for r in results
        ]
    )


def initialize_eval_interface(example: dict, model_outputs: list[dict]):
    """Initialize the interface with example text."""
    try:
        html_content = create_bonus_html(example["leadin"], example["parts"])

        # Create confidence plot data
        plot_data = create_bonus_confidence_plot(example["parts"], model_outputs)

        # Store state
        state = json.dumps({"parts": example["parts"], "outputs": model_outputs})

        return html_content, plot_data, state
    except Exception as e:
        logging.error(f"Error initializing interface: {e}", exc_info=True)
        return f"<div>Error initializing interface: {str(e)}</div>", pd.DataFrame(), "{}"


def validate_workflow(workflow: Workflow):
    """Validate that a workflow is properly configured for the bonus task."""
    if not workflow.steps:
        raise ValueError("Workflow must have at least one step")

    # Ensure all steps are properly configured
    for step_id, step in workflow.steps.items():
        validate_model_step(step)

    # Check that the workflow has the correct structure
    input_vars = set(workflow.inputs)
    if "leadin" not in input_vars or "part" not in input_vars:
        raise ValueError("Workflow must have 'leadin' and 'part' as inputs")

    output_vars = set(workflow.outputs)
    if not all(var in output_vars for var in ["answer", "confidence", "explanation"]):
        raise ValueError("Workflow must produce 'answer', 'confidence', and 'explanation' as outputs")


def validate_model_step(model_step: ModelStep):
    """Validate that a model step is properly configured for the bonus task."""
    # Check required fields
    if not model_step.model or not model_step.provider:
        raise ValueError("Model step must have both model and provider specified")

    if model_step.call_type != "llm":
        raise ValueError("Model step must have call_type 'llm'")

    # Validate temperature for LLM steps
    if model_step.temperature is None:
        raise ValueError("Temperature must be specified for LLM model steps")

    if not (0.0 <= model_step.temperature <= 1.0):
        raise ValueError(f"Temperature must be between 0.0 and 1.0, got {model_step.temperature}")

    # Validate input fields
    input_field_names = {field.name for field in model_step.input_fields}
    if "leadin" not in input_field_names or "part" not in input_field_names:
        raise ValueError("Model step must have 'leadin' and 'part' input fields")

    # Validate output fields
    output_field_names = {field.name for field in model_step.output_fields}
    required_outputs = {"answer", "confidence", "explanation"}
    if not all(out in output_field_names for out in required_outputs):
        raise ValueError("Model step must have all required output fields: answer, confidence, explanation")

    # Validate confidence output field is of type float
    for field in model_step.output_fields:
        if field.name == "confidence" and field.type != "float":
            raise ValueError("The 'confidence' output field must be of type 'float'")


class BonusInterface:
    """Gradio interface for the Bonus mode."""

    def __init__(self, app: gr.Blocks, dataset: Dataset, model_options: dict, defaults: dict):
        """Initialize the Bonus interface."""
        logging.info(f"Initializing Bonus interface with dataset size: {len(dataset)}")
        self.ds = dataset
        self.model_options = model_options
        self.app = app
        self.defaults = defaults
        self.output_state = gr.State(value="{}")
        self.render()

    def _render_model_interface(self, workflow: Workflow, simple: bool = True):
        """Render the model interface."""
        self.pipeline_interface = PipelineInterface(
            workflow,
            simple=simple,
            model_options=list(self.model_options.keys()),
        )

    def _render_qb_interface(self):
        """Render the quizbowl interface."""
        with gr.Row(elem_classes="bonus-header-row form-inline"):
            self.qid_selector = get_qid_selector(len(self.ds))
            self.run_btn = gr.Button("Run on Bonus Question", variant="secondary")

        self.question_display = gr.HTML(label="Question", elem_id="bonus-question-display")
        with gr.Row():
            self.confidence_plot = gr.Plot(
                label="Part Confidence",
                format="webp",
            )

        self.results_table = gr.DataFrame(
            label="Model Outputs",
            value=pd.DataFrame(columns=["Part", "Correct?", "Confidence", "Prediction", "Explanation"]),
        )

        with gr.Row():
            self.eval_btn = gr.Button("Evaluate", variant="primary")

        with gr.Accordion("Model Submission", elem_classes="model-submission-accordion", open=True):
            with gr.Row():
                self.model_name_input = gr.Textbox(label="Model Name")
                self.description_input = gr.Textbox(label="Description")
            with gr.Row():
                gr.LoginButton()
                self.submit_btn = gr.Button("Submit", variant="primary")
            self.submit_status = gr.HTML(label="Submission Status")

    def render(self):
        """Create the Gradio interface."""
        self.hidden_input = gr.Textbox(value="", visible=False, elem_id="hidden-index")
        workflow = self.defaults["init_workflow"]

        with gr.Row():
            # Model Panel
            with gr.Column(scale=1):
                self._render_model_interface(workflow, simple=self.defaults["simple_workflow"])

            with gr.Column(scale=1):
                self._render_qb_interface()

        self._setup_event_listeners()

    def get_new_question_html(self, question_id: int):
        """Get the HTML for a new question."""
        if question_id is None:
            logging.error("Question ID is None. Setting to 1")
            question_id = 1
        try:
            question_id = int(question_id) - 1
            if not self.ds or question_id < 0 or question_id >= len(self.ds):
                return "Invalid question ID or dataset not loaded"

            example = self.ds[question_id]
            leadin = example["leadin"]
            parts = example["parts"]
            return create_bonus_html(leadin, parts)
        except Exception as e:
            return f"Error loading question: {str(e)}"

    def get_model_outputs(self, example: dict, pipeline_state: PipelineState):
        """Get the model outputs for a given question ID."""
        outputs = []
        leadin = example["leadin"]
        workflow = pipeline_state.workflow
        if len(workflow.steps) > 1:
            agent = MultiStepBonusAgent(workflow)
        else:
            agent = SimpleBonusAgent(workflow)

        for i, part in enumerate(example["parts"]):
            # Run model for each part
            part_output = agent.run(leadin, part["part"])

            # Add part number and evaluate score
            part_output["part_number"] = i + 1
            part_output["score"] = evaluate_prediction(part_output["answer"], part["clean_answers"])

            outputs.append(part_output)

        return outputs

    def single_run(
        self,
        question_id: int,
        pipeline_state: PipelineState,
    ) -> tuple[str, Any, Any]:
        """Run the agent in bonus mode."""
        try:
            # Validate inputs
            question_id = int(question_id - 1)
            if not self.ds or question_id < 0 or question_id >= len(self.ds):
                return "Invalid question ID or dataset not loaded", None, None

            example = self.ds[question_id]
            outputs = self.get_model_outputs(example, pipeline_state)

            # Process results and prepare visualization data
            html_content, plot_data, output_state = initialize_eval_interface(example, outputs)
            df = process_bonus_results(outputs)

            return (
                html_content,
                gr.update(value=plot_data, label=f"Part Confidence on Question {question_id + 1}"),
                gr.update(value=output_state),
                gr.update(value=df, label=f"Model Outputs for Question {question_id + 1}"),
            )
        except Exception as e:
            import traceback

            error_msg = f"Error: {str(e)}\n{traceback.format_exc()}"
            return error_msg, gr.skip(), gr.skip(), gr.skip()

    def evaluate(self, pipeline_state: PipelineState, progress: gr.Progress = gr.Progress()):
        """Evaluate the bonus questions."""
        try:
            # Validate inputs
            if not self.ds or not self.ds.num_rows:
                return "No dataset loaded", None, None

            total_correct = 0
            total_parts = 0
            part_scores = []
            part_numbers = []

            for example in progress.tqdm(self.ds, desc="Evaluating bonus questions"):
                model_outputs = self.get_model_outputs(example, pipeline_state)

                for output in model_outputs:
                    total_parts += 1
                    if output["score"] == 1:
                        total_correct += 1
                    part_scores.append(output["score"])
                    part_numbers.append(output["part_number"])

            accuracy = total_correct / total_parts
            df = pd.DataFrame(
                [
                    {
                        "Part Accuracy": f"{accuracy:.2%}",
                        "Total Score": f"{total_correct}/{total_parts}",
                        "Questions Evaluated": len(self.ds),
                    }
                ]
            )

            plot_data = create_scatter_pyplot(part_numbers, part_scores)
            return (
                gr.update(value=df, label="Scores on Sample Set"),
                gr.update(value=plot_data, label="Part Scores on Sample Set"),
            )
        except Exception:
            import traceback

            logging.error(f"Error evaluating bonus: {traceback.format_exc()}")
            return "Error evaluating bonus", None, None

    def submit_model(
        self, model_name: str, description: str, pipeline_state: PipelineState, profile: gr.OAuthProfile = None
    ):
        """Submit the model output."""
        return submit.submit_model(model_name, description, pipeline_state.workflow, "bonus", profile)

    def _setup_event_listeners(self):
        # Initialize with the default question (ID 0)

        gr.on(
            triggers=[self.app.load, self.qid_selector.change],
            fn=self.get_new_question_html,
            inputs=[self.qid_selector],
            outputs=[self.question_display],
        )

        self.run_btn.click(
            self.pipeline_interface.validate_workflow,
            inputs=[self.pipeline_interface.pipeline_state],
            outputs=[self.pipeline_interface.pipeline_state],
        ).success(
            self.single_run,
            inputs=[
                self.qid_selector,
                self.pipeline_interface.pipeline_state,
            ],
            outputs=[
                self.question_display,
                self.confidence_plot,
                self.output_state,
                self.results_table,
            ],
        )

        self.eval_btn.click(
            fn=self.evaluate,
            inputs=[self.pipeline_interface.pipeline_state],
            outputs=[self.results_table, self.confidence_plot],
        )

        self.submit_btn.click(
            fn=self.submit_model,
            inputs=[
                self.model_name_input,
                self.description_input,
                self.pipeline_interface.pipeline_state,
            ],
            outputs=[self.submit_status],
        )
        self.hidden_input.change(
            fn=update_tossup_plot,
            inputs=[self.hidden_input, self.output_state],
            outputs=[self.confidence_plot],
        )