File size: 18,992 Bytes
193db9d d15e788 193db9d 02b7dec 97fcd0c d15e788 0bab47c 9756440 d15e788 193db9d 0bab47c 193db9d 97fcd0c 193db9d 3b39b49 193db9d 3b39b49 193db9d 973519b 193db9d 0bab47c 193db9d 3b39b49 193db9d d15e788 193db9d 3b39b49 193db9d 3b39b49 193db9d 3b39b49 193db9d d15e788 193db9d 0bab47c 193db9d 0bab47c 193db9d 9756440 193db9d d15e788 193db9d 0bab47c 193db9d 02b7dec 0bab47c 9756440 193db9d 0bab47c 193db9d 3b39b49 d15e788 0bab47c 3b39b49 0bab47c 193db9d 02b7dec 193db9d 02b7dec 193db9d 3b39b49 193db9d 3b39b49 193db9d 0bab47c 193db9d 9756440 193db9d 9756440 193db9d 3b39b49 193db9d 3b39b49 d15e788 3b39b49 193db9d 0bab47c 193db9d 0bab47c 193db9d 02b7dec d15e788 9756440 02b7dec 9756440 02b7dec 9756440 02b7dec 9756440 d15e788 3b39b49 193db9d 9756440 193db9d 9756440 0bab47c 193db9d 0bab47c 193db9d 0bab47c 02b7dec 0bab47c 193db9d 0bab47c 02b7dec 0bab47c 193db9d 9756440 3b39b49 193db9d 9756440 193db9d 0bab47c 193db9d 02b7dec 0bab47c 193db9d d15e788 0bab47c d15e788 0bab47c 02b7dec 0bab47c 193db9d 9756440 193db9d 9756440 193db9d 3b39b49 193db9d d15e788 02b7dec d15e788 9756440 02b7dec 9756440 02b7dec 9756440 d15e788 193db9d 9756440 193db9d 3b39b49 193db9d 0bab47c 193db9d 0bab47c 193db9d 3b39b49 0bab47c 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 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 411 412 413 414 415 416 417 418 419 420 421 422 423 424 425 426 427 428 429 430 431 432 433 434 435 436 437 438 439 440 441 442 443 444 445 446 447 448 449 450 451 452 453 454 455 456 457 458 459 460 461 462 463 464 |
import json
from typing import Any
import gradio as gr
import numpy as np
import pandas as pd
from datasets import Dataset
from loguru import logger
from app_configs import UNSELECTED_PIPELINE_NAME
from components import commons
from components.model_pipeline.model_pipeline import PipelineInterface, PipelineState, PipelineUIState
from components.model_pipeline.tossup_pipeline import TossupPipelineInterface, TossupPipelineState
from components.typed_dicts import PipelineStateDict, TossupPipelineStateDict
from display.formatting import styled_error
from submission import submit
from workflows.qb_agents import QuizBowlTossupAgent, TossupResult
from workflows.structs import ModelStep, TossupWorkflow
from . import populate
from .plotting import (
create_scatter_pyplot,
create_tossup_confidence_pyplot,
create_tossup_html,
update_tossup_plot,
)
from .utils import evaluate_prediction
# TODO: Error handling on run tossup and evaluate tossup and show correct messages
# TODO: ^^ Same for Bonus
class ScoredTossupResult(TossupResult):
"""Result of a tossup question with evaluation score and position."""
score: int # Correctness score of the answer
token_position: int # Position in the question where prediction was made
def add_model_scores(model_outputs: list[dict], clean_answers: list[str], run_indices: list[int]) -> list[dict]:
"""Add model scores to the model outputs."""
for output, run_idx in zip(model_outputs, run_indices):
output["score"] = evaluate_prediction(output["answer"], clean_answers)
output["token_position"] = run_idx + 1
return model_outputs
def prepare_buzz_evals(
run_indices: list[int], model_outputs: list[dict]
) -> tuple[list[str], list[tuple[int, float, bool]]]:
"""Process text into tokens and assign random values for demonstration."""
if not run_indices:
logger.warning("No run indices provided, returning empty results")
return [], []
eval_points = []
for i, v in zip(run_indices, model_outputs):
eval_points.append((int(i), v))
return eval_points
def initialize_eval_interface(example, model_outputs: list[dict]):
"""Initialize the interface with example text."""
try:
tokens = example["question"].split()
run_indices = example["run_indices"]
answer = example["answer_primary"]
clean_answers = example["clean_answers"]
eval_points = prepare_buzz_evals(run_indices, model_outputs)
if not tokens:
return "<div>No tokens found in the provided text.</div>", pd.DataFrame(), "{}"
highlighted_index = next((int(i) for i, v in eval_points if v["buzz"] == 1), -1)
html_content = create_tossup_html(tokens, answer, clean_answers, run_indices, eval_points)
plot_data = create_tossup_confidence_pyplot(tokens, eval_points, highlighted_index)
# Store tokens, values, and buzzes as JSON for later use
state = json.dumps({"tokens": tokens, "values": eval_points})
return html_content, plot_data, state
except Exception as e:
logger.exception(f"Error initializing interface: {e.args}")
return f"<div>Error initializing interface: {str(e)}</div>", pd.DataFrame(), "{}"
def process_tossup_results(results: list[dict], top_k_mode: bool = False) -> pd.DataFrame:
"""Process results from tossup mode and prepare visualization data."""
# Create DataFrame for detailed results
if top_k_mode:
raise ValueError("Top-k mode not supported for tossup mode")
return pd.DataFrame(
[
{
"Token Position": r["token_position"],
"Correct?": "✅" if r["score"] == 1 else "❌",
"Confidence": r["confidence"],
"Prediction": r["answer"],
}
for r in results
]
)
def validate_workflow(workflow: TossupWorkflow):
"""
Validate that a workflow is properly configured for the tossup task.
Args:
workflow (TossupWorkflow): The workflow to validate
Raises:
ValueError: If the workflow is not properly configured
"""
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 "question" not in input_vars:
raise ValueError("Workflow must have 'question' as an input")
output_vars = set(workflow.outputs)
if not any("answer" in out_var for out_var in output_vars):
raise ValueError("Workflow must produce an 'answer' as output")
if not any("confidence" in out_var for out_var in output_vars):
raise ValueError("Workflow must produce a 'confidence' score as output")
def validate_model_step(model_step: ModelStep):
"""
Validate that a model step is properly configured for the tossup task.
Args:
model_step (ModelStep): The model step to validate
Raises:
ValueError: If the model step is not properly configured
"""
# 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 "question" not in input_field_names:
raise ValueError("Model step must have a 'question' input field")
# Validate output fields
output_field_names = {field.name for field in model_step.output_fields}
if "answer" not in output_field_names:
raise ValueError("Model step must have an 'answer' output field")
if "confidence" not in output_field_names:
raise ValueError("Model step must have a 'confidence' output field")
# 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 TossupInterface:
"""Gradio interface for the Tossup mode."""
def __init__(self, app: gr.Blocks, dataset: Dataset, model_options: dict, defaults: dict):
"""Initialize the Tossup interface."""
logger.info(f"Initializing Tossup 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_pipeline_interface(self, workflow: TossupWorkflow, simple: bool = True):
"""Render the model interface."""
with gr.Row(elem_classes="bonus-header-row form-inline"):
self.pipeline_selector = commons.get_pipeline_selector([])
self.load_btn = gr.Button("⬇️ Import Pipeline", variant="secondary")
self.pipeline_interface = TossupPipelineInterface(
self.app,
workflow,
simple=simple,
model_options=list(self.model_options.keys()),
defaults=self.defaults,
)
def _render_qb_interface(self):
"""Render the quizbowl interface."""
with gr.Row(elem_classes="bonus-header-row form-inline"):
self.qid_selector = commons.get_qid_selector(len(self.ds))
self.early_stop_checkbox = gr.Checkbox(
value=self.defaults["early_stop"],
label="Early Stop",
info="Stop if already buzzed",
scale=0,
)
self.run_btn = gr.Button("Run on Tossup Question", variant="secondary")
self.question_display = gr.HTML(label="Question", elem_id="tossup-question-display")
self.error_display = gr.HTML(label="Error", elem_id="tossup-error-display", visible=False)
with gr.Row():
self.confidence_plot = gr.Plot(
label="Buzz Confidence",
format="webp",
)
self.model_outputs_display = gr.JSON(label="Model Outputs", value="{}", show_indices=True, visible=False)
self.results_table = gr.DataFrame(
label="Model Outputs",
value=pd.DataFrame(columns=["Token Position", "Correct?", "Confidence", "Prediction"]),
visible=False,
)
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_pipeline_interface(workflow, simple=self.defaults["simple_workflow"])
with gr.Column(scale=1):
self._render_qb_interface()
self._setup_event_listeners()
def validate_workflow(self, state_dict: TossupPipelineStateDict):
"""Validate the workflow."""
try:
pipeline_state = TossupPipelineState(**state_dict)
validate_workflow(pipeline_state.workflow)
except Exception as e:
raise gr.Error(f"Error validating workflow: {str(e)}")
def get_new_question_html(self, question_id: int) -> str:
"""Get the HTML for a new question."""
if question_id is None:
logger.error("Question ID is None. Setting to 1")
question_id = 1
try:
example = self.ds[question_id - 1]
question_tokens = example["question"].split()
return create_tossup_html(
question_tokens, example["answer_primary"], example["clean_answers"], example["run_indices"]
)
except Exception as e:
return f"Error loading question: {str(e)}"
def get_model_outputs(
self, example: dict, pipeline_state: PipelineState, early_stop: bool
) -> list[ScoredTossupResult]:
"""Get the model outputs for a given question ID."""
question_runs = []
tokens = example["question"].split()
for run_idx in example["run_indices"]:
question_runs.append(" ".join(tokens[: run_idx + 1]))
agent = QuizBowlTossupAgent(pipeline_state.workflow)
outputs = list(agent.run(question_runs, early_stop=early_stop))
outputs = add_model_scores(outputs, example["clean_answers"], example["run_indices"])
return outputs
def get_pipeline_names(self, profile: gr.OAuthProfile | None) -> list[str]:
names = [UNSELECTED_PIPELINE_NAME] + populate.get_pipeline_names("tossup", profile)
return gr.update(choices=names, value=UNSELECTED_PIPELINE_NAME)
def load_pipeline(
self, model_name: str, pipeline_change: bool, profile: gr.OAuthProfile | None
) -> tuple[str, PipelineStateDict, bool, dict]:
try:
workflow = populate.load_workflow("tossup", model_name, profile)
if workflow is None:
logger.warning(f"Could not load workflow for {model_name}")
return UNSELECTED_PIPELINE_NAME, gr.skip(), gr.skip(), gr.update(visible=False)
pipeline_state_dict = TossupPipelineState.from_workflow(workflow).model_dump()
return UNSELECTED_PIPELINE_NAME, pipeline_state_dict, not pipeline_change, gr.update(visible=True)
except Exception as e:
logger.exception(e)
error_msg = styled_error(f"Error loading pipeline: {str(e)}")
return UNSELECTED_PIPELINE_NAME, gr.skip(), gr.skip(), gr.update(visible=True, value=error_msg)
def single_run(
self,
question_id: int,
state_dict: TossupPipelineStateDict,
early_stop: bool = True,
) -> tuple[str, Any, Any]:
"""Run the agent in tossup mode with a system prompt."""
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]
pipeline_state = TossupPipelineState(**state_dict)
outputs = self.get_model_outputs(example, pipeline_state, early_stop)
# Process results and prepare visualization data
tokens_html, plot_data, output_state = initialize_eval_interface(example, outputs)
df = process_tossup_results(outputs)
step_outputs = [output["step_outputs"] for output in outputs]
return (
tokens_html,
gr.update(value=output_state),
gr.update(value=plot_data, label=f"Buzz Confidence on Question {question_id + 1}"),
gr.update(value=df, label=f"Model Outputs for Question {question_id + 1}", visible=True),
gr.update(value=step_outputs, label=f"Step Outputs for Question {question_id + 1}", visible=True),
gr.update(visible=False),
)
except Exception as e:
import traceback
error_msg = styled_error(f"Error: {str(e)}\n{traceback.format_exc()}")
return (
gr.skip(),
gr.skip(),
gr.skip(),
gr.update(visible=False),
gr.update(visible=False),
gr.update(visible=True, value=error_msg),
)
def evaluate(self, state_dict: TossupPipelineStateDict, progress: gr.Progress = gr.Progress()):
"""Evaluate the tossup questions."""
try:
# Validate inputs
if not self.ds or not self.ds.num_rows:
return "No dataset loaded", None, None
pipeline_state = TossupPipelineState(**state_dict)
buzz_counts = 0
correct_buzzes = 0
token_positions = []
correctness = []
for example in progress.tqdm(self.ds, desc="Evaluating tossup questions"):
model_outputs = self.get_model_outputs(example, pipeline_state, early_stop=True)
if model_outputs[-1]["buzz"]:
buzz_counts += 1
if model_outputs[-1]["score"] == 1:
correct_buzzes += 1
token_positions.append(model_outputs[-1]["token_position"])
correctness.append(model_outputs[-1]["score"])
buzz_accuracy = correct_buzzes / buzz_counts
df = pd.DataFrame(
[
{
"Avg Buzz Position": f"{np.mean(token_positions):.2f}",
"Buzz Accuracy": f"{buzz_accuracy:.2%}",
"Total Score": f"{correct_buzzes}/{len(self.ds)}",
}
]
)
plot_data = create_scatter_pyplot(token_positions, correctness)
return (
gr.update(value=plot_data, label="Buzz Positions on Sample Set"),
gr.update(value=df, label="Scores on Sample Set", visible=True),
gr.update(visible=False),
)
except Exception as e:
import traceback
logger.exception(f"Error evaluating tossups: {e.args}")
return (
gr.skip(),
gr.update(visible=False),
gr.update(visible=True, value=styled_error(f"Error: {str(e)}\n{traceback.format_exc()}")),
)
def submit_model(
self,
model_name: str,
description: str,
state_dict: TossupPipelineStateDict,
profile: gr.OAuthProfile = None,
):
"""Submit the model output."""
pipeline_state = TossupPipelineState(**state_dict)
return submit.submit_model(model_name, description, pipeline_state.workflow, "tossup", profile)
def _setup_event_listeners(self):
gr.on(
triggers=[self.app.load, self.qid_selector.change],
fn=self.get_new_question_html,
inputs=[self.qid_selector],
outputs=[self.question_display],
)
gr.on(
triggers=[self.app.load],
fn=self.get_pipeline_names,
outputs=[self.pipeline_selector],
)
pipeline_state = self.pipeline_interface.pipeline_state
pipeline_change = self.pipeline_interface.pipeline_change
self.load_btn.click(
fn=self.load_pipeline,
inputs=[self.pipeline_selector, pipeline_change],
outputs=[self.pipeline_selector, pipeline_state, pipeline_change, self.error_display],
)
self.pipeline_interface.add_triggers_for_pipeline_export([pipeline_state.change], pipeline_state)
self.run_btn.click(
self.pipeline_interface.validate_workflow,
inputs=[self.pipeline_interface.pipeline_state],
outputs=[],
).success(
self.single_run,
inputs=[
self.qid_selector,
self.pipeline_interface.pipeline_state,
self.early_stop_checkbox,
],
outputs=[
self.question_display,
self.output_state,
self.confidence_plot,
self.results_table,
self.model_outputs_display,
self.error_display,
],
)
self.eval_btn.click(
fn=self.evaluate,
inputs=[self.pipeline_interface.pipeline_state],
outputs=[self.confidence_plot, self.results_table, self.error_display],
)
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],
)
|