Skip to content

AutoLabeler

Source code in autolabel/src/autolabel/labeler.py
 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
465
466
467
468
469
470
471
472
473
474
475
476
477
478
479
480
481
482
483
484
485
486
487
488
489
490
491
492
493
494
495
496
497
498
499
500
501
502
503
504
505
506
507
508
509
510
511
512
513
514
515
516
517
518
519
520
521
522
523
524
525
526
527
528
529
530
531
532
533
534
535
536
537
538
539
540
541
542
543
544
545
546
547
548
549
550
551
552
553
554
555
556
557
558
559
560
561
562
563
564
565
566
567
568
569
570
571
572
573
574
575
576
577
578
579
580
581
582
583
584
585
586
587
588
589
590
591
592
593
594
595
596
597
598
599
600
601
602
603
604
605
606
607
608
609
610
611
612
613
614
615
616
617
618
619
620
621
622
623
624
625
626
627
628
629
630
631
632
633
634
635
636
637
638
639
640
641
642
643
644
645
646
647
648
649
650
651
652
653
654
655
656
657
658
659
660
661
662
663
664
665
666
667
668
669
670
671
672
673
674
675
676
677
678
679
680
681
682
683
684
685
686
687
688
689
690
691
692
693
694
695
696
697
698
699
700
701
702
703
704
705
706
707
708
709
710
711
712
713
714
715
716
717
718
719
720
721
722
723
724
class LabelingAgent:
    COST_KEY = "Cost in $"
    CONFIDENCE_MAX_CONTEXT_LENGTH = 3400

    def __init__(
        self,
        config: Union[AutolabelConfig, str, dict],
        cache: Optional[bool] = True,
        example_selector: Optional[BaseExampleSelector] = None,
        console_output: Optional[bool] = True,
        generation_cache: Optional[BaseCache] = SQLAlchemyGenerationCache(),
        transform_cache: Optional[BaseCache] = SQLAlchemyTransformCache(),
        confidence_cache: Optional[BaseCache] = SQLAlchemyConfidenceCache(),
        confidence_tokenizer: Optional[AutoTokenizer] = None,
    ) -> None:
        self.generation_cache = generation_cache
        self.transform_cache = transform_cache
        self.confidence_cache = confidence_cache
        if not cache:
            logger.warning(
                f"cache parameter is deprecated and will be removed soon. Please use generation_cache, transform_cache and confidence_cache instead."
            )
            self.generation_cache = None
            self.transform_cache = None
            self.confidence_cache = None

        if self.generation_cache is not None:
            self.generation_cache.initialize()
        if self.transform_cache is not None:
            self.transform_cache.initialize()
        if self.confidence_cache is not None:
            self.confidence_cache.initialize()

        self.console = Console(quiet=not console_output)

        self.config = (
            config if isinstance(config, AutolabelConfig) else AutolabelConfig(config)
        )
        self.task = TaskFactory.from_config(self.config)
        self.llm: BaseModel = ModelFactory.from_config(
            self.config, cache=self.generation_cache
        )

        if self.config.confidence_chunk_column():
            if not confidence_tokenizer:
                self.confidence_tokenizer = AutoTokenizer.from_pretrained(
                    "google/flan-t5-xxl"
                )
            else:
                self.confidence_tokenizer = confidence_tokenizer
        score_type = "logprob_average"
        if self.config.task_type() == TaskType.ATTRIBUTE_EXTRACTION:
            score_type = "logprob_average_per_key"
        self.confidence = ConfidenceCalculator(
            score_type=score_type,
            llm=self.llm,
            cache=self.confidence_cache,
        )

        self.example_selector = example_selector

        if in_notebook():
            import nest_asyncio

            nest_asyncio.apply()

    def run(
        self,
        dataset: AutolabelDataset,
        output_name: Optional[str] = None,
        max_items: Optional[int] = None,
        start_index: int = 0,
        additional_metrics: Optional[List[BaseMetric]] = [],
        skip_eval: Optional[bool] = False,
    ) -> Tuple[pd.Series, pd.DataFrame, List[MetricResult]]:
        return asyncio.run(
            self.arun(
                dataset=dataset,
                output_name=output_name,
                max_items=max_items,
                start_index=start_index,
                additional_metrics=additional_metrics,
                skip_eval=skip_eval,
            )
        )

    async def arun(
        self,
        dataset: AutolabelDataset,
        output_name: Optional[str] = None,
        max_items: Optional[int] = None,
        start_index: int = 0,
        additional_metrics: Optional[List[BaseMetric]] = [],
        skip_eval: Optional[bool] = False,
    ) -> Tuple[pd.Series, pd.DataFrame, List[MetricResult]]:
        """Labels data in a given dataset. Output written to new CSV file.

        Args:
            dataset: path to CSV dataset to be annotated
            max_items: maximum items in dataset to be annotated
            output_name: custom name of output CSV file
            start_index: skips annotating [0, start_index)
        """

        dataset = dataset.get_slice(max_items=max_items, start_index=start_index)

        llm_labels = []

        # Get the seed examples from the dataset config
        seed_examples = self.config.few_shot_example_set()

        # If this dataset config is a string, read the corrresponding csv file
        if isinstance(seed_examples, str):
            seed_loader = AutolabelDataset(seed_examples, self.config)
            seed_examples = seed_loader.inputs

        # Check explanations are present in data if explanation_column is passed in
        if (
            self.config.explanation_column()
            and len(seed_examples) > 0
            and self.config.explanation_column() not in list(seed_examples[0].keys())
        ):
            raise ValueError(
                f"Explanation column {self.config.explanation_column()} not found in dataset.\nMake sure that explanations were generated using labeler.generate_explanations(seed_file)."
            )

        if self.example_selector is None:
            if (
                self.config.label_selection()
                and self.config.few_shot_algorithm() != "fixed"
            ):
                # TODO: Add support for other few shot algorithms specially semantic similarity
                raise ValueError(
                    "Error: Only 'fixed' few shot example selector is supported for label selection."
                )

            self.example_selector = ExampleSelectorFactory.initialize_selector(
                self.config,
                [safe_serialize_to_string(example) for example in seed_examples],
                dataset.df.keys().tolist(),
                cache=self.generation_cache is not None,
            )

        if self.config.label_selection():
            if self.config.task_type() != TaskType.CLASSIFICATION:
                self.console.print(
                    "Warning: label_selection only supported for classification tasks!"
                )
            else:
                self.label_selector = LabelSelector(
                    config=self.config,
                    embedding_func=PROVIDER_TO_MODEL.get(
                        self.config.embedding_provider(), DEFAULT_EMBEDDING_PROVIDER
                    )(model=self.config.embedding_model_name()),
                )

        current_index = 0
        cost = 0.0
        postfix_dict = {}

        indices = range(current_index, len(dataset.inputs))
        selected_labels = self.config.labels_list()
        for current_index in track_with_stats(
            indices,
            postfix_dict,
            total=len(dataset.inputs) - current_index,
            console=self.console,
        ):
            chunk = dataset.inputs[current_index]
            examples = []

            if (
                self.config.label_selection()
                and self.config.task_type() == TaskType.CLASSIFICATION
            ):
                # get every column except the one we want to label
                toEmbed = chunk.copy()
                if self.config.label_column() and self.config.label_column() in toEmbed:
                    del toEmbed[self.config.label_column()]

                # convert this to a string
                toEmbed = json.dumps(toEmbed)

                selected_labels = self.label_selector.select_labels(toEmbed)

                if self.example_selector:
                    examples = self.example_selector.select_examples(
                        safe_serialize_to_string(chunk),
                        selected_labels=selected_labels,
                        label_column=self.config.label_column(),
                    )
                else:
                    examples = []
            else:
                if self.example_selector:
                    examples = self.example_selector.select_examples(
                        safe_serialize_to_string(chunk),
                    )

            # Construct Prompt to pass to LLM
            final_prompt = self.task.construct_prompt(
                chunk,
                examples,
                selected_labels=selected_labels,
                max_input_tokens=self.llm.DEFAULT_CONTEXT_LENGTH,
                get_num_tokens=self.llm.get_num_tokens,
            )

            response = await self.llm.label([final_prompt])
            for i, generations, error, latency in zip(
                range(len(response.generations)),
                response.generations,
                response.errors,
                response.latencies,
            ):
                input_tokens = self.llm.get_num_tokens(final_prompt)
                if error is not None:
                    annotation = LLMAnnotation(
                        successfully_labeled=False,
                        label=self.task.NULL_LABEL_TOKEN,
                        raw_response="",
                        curr_sample=pickle.dumps(chunk),
                        prompt=final_prompt,
                        confidence_score=0,
                        error=error,
                        input_tokens=input_tokens,
                        cost=0,
                        latency=0,
                    )
                else:
                    annotations = []
                    for generation in generations:
                        annotation = self.task.parse_llm_response(
                            generation, chunk, final_prompt
                        )
                        annotation.confidence_prompt = (
                            self.task.construct_confidence_prompt(chunk, examples)
                        )
                        annotation.input_tokens = input_tokens
                        annotation.output_tokens = self.llm.get_num_tokens(
                            annotation.raw_response
                        )
                        annotation.cost = sum(response.costs)
                        annotation.latency = latency

                        if self.config.confidence():
                            try:
                                annotation.confidence_score = (
                                    await self.get_confidence_score(
                                        annotation, chunk, examples
                                    )
                                )
                            except Exception as e:
                                logger.exception(
                                    f"Error calculating confidence score: {e}"
                                )
                                logger.warning(
                                    f"Could not calculate confidence score for annotation: {annotation}"
                                )
                                if (
                                    self.config.task_type()
                                    == TaskType.ATTRIBUTE_EXTRACTION
                                ):
                                    annotation.confidence_score = {}
                                else:
                                    annotation.confidence_score = 0

                        annotations.append(annotation)
                    annotation = self.majority_annotation(annotations)

                llm_labels.append(annotation)

            cost += sum(response.costs)
            postfix_dict[self.COST_KEY] = f"{cost:.2f}"

            # Evaluate the task every eval_every examples
            if not skip_eval and (current_index + 1) % 100 == 0:
                if dataset.gt_labels:
                    eval_result = self.task.eval(
                        llm_labels,
                        (
                            dataset.gt_labels[: len(llm_labels)]
                            if isinstance(dataset.gt_labels, list)
                            else {
                                k: v[: len(llm_labels)]
                                for k, v in dataset.gt_labels.items()
                            }
                        ),
                        additional_metrics=additional_metrics,
                    )

                    for m in eval_result:
                        # This is a row wise metric
                        if isinstance(m.value, list):
                            continue
                        elif m.show_running:
                            postfix_dict[m.name] = (
                                f"{m.value:.4f}"
                                if isinstance(m.value, float)
                                else m.value
                            )

        eval_result = None
        table = {}

        # if true labels are provided, evaluate accuracy of predictions
        if not skip_eval and dataset.gt_labels:
            eval_result = self.task.eval(
                llm_labels,
                (
                    dataset.gt_labels[: len(llm_labels)]
                    if isinstance(dataset.gt_labels, list)
                    else {k: v[: len(llm_labels)] for k, v in dataset.gt_labels.items()}
                ),
                additional_metrics=additional_metrics,
            )
            # TODO: serialize and write to file
            for m in eval_result:
                if isinstance(m.value, list):
                    continue
                elif m.show_running:
                    table[m.name] = m.value
                else:
                    self.console.print(f"{m.name}:\n{m.value}")

        # print cost
        self.console.print(f"Actual Cost: {maybe_round(cost)}")
        print_table(table, console=self.console, default_style=METRIC_TABLE_STYLE)

        dataset.process_labels(llm_labels, eval_result)
        # Only save to csv if output_name is provided or dataset is a string
        if not output_name and isinstance(dataset, str):
            output_name = (
                dataset.rsplit(".", 1)[0] + "_labeled." + dataset.rsplit(".", 1)[1]
            )

        if output_name:
            dataset.save(output_file_name=output_name)
        return dataset

    def plan(
        self,
        dataset: AutolabelDataset,
        max_items: Optional[int] = None,
        start_index: int = 0,
    ) -> None:
        """Calculates and prints the cost of calling autolabel.run() on a given dataset

        Args:
            dataset: path to a CSV dataset
        """
        dataset = dataset.get_slice(max_items=max_items, start_index=start_index)

        if (
            self.config.confidence()
            and "REFUEL_API_KEY" not in os.environ
            and not self.llm.returns_token_probs()
        ):
            raise ValueError(
                "REFUEL_API_KEY environment variable must be set to compute confidence scores. You can request an API key at https://refuel-ai.typeform.com/llm-access."
            )

        prompt_list = []
        total_cost = 0

        # Get the seed examples from the dataset config
        seed_examples = self.config.few_shot_example_set()

        # If this dataset config is a string, read the corrresponding csv file
        if isinstance(seed_examples, str):
            seed_loader = AutolabelDataset(seed_examples, self.config)
            seed_examples = seed_loader.inputs

        # Check explanations are present in data if explanation_column is passed in
        if (
            self.config.explanation_column()
            and len(seed_examples) > 0
            and self.config.explanation_column() not in list(seed_examples[0].keys())
        ):
            raise ValueError(
                f"Explanation column {self.config.explanation_column()} not found in dataset.\nMake sure that explanations were generated using labeler.generate_explanations(seed_file)."
            )

        self.example_selector = ExampleSelectorFactory.initialize_selector(
            self.config,
            [safe_serialize_to_string(example) for example in seed_examples],
            dataset.df.keys().tolist(),
            cache=self.generation_cache is not None,
        )

        if self.config.label_selection():
            if self.config.task_type() != TaskType.CLASSIFICATION:
                self.console.print(
                    "Warning: label_selection only supported for classification tasks!"
                )
            else:
                self.label_selector = LabelSelector(
                    config=self.config,
                    embedding_func=PROVIDER_TO_MODEL.get(
                        self.config.embedding_provider(), DEFAULT_EMBEDDING_PROVIDER
                    )(model=self.config.embedding_model_name()),
                )

        input_limit = min(len(dataset.inputs), 100) if max_items is None else max_items  # type: ignore
        for input_i in track(
            dataset.inputs[:input_limit],
            description="Generating Prompts...",
            console=self.console,
        ):
            # TODO: Check if this needs to use the example selector
            if self.example_selector:
                examples = self.example_selector.select_examples(
                    safe_serialize_to_string(input_i)
                )
            else:
                examples = []
            if (
                self.config.label_selection()
                and self.config.task_type() == TaskType.CLASSIFICATION
            ):
                selected_labels = self.label_selector.select_labels(input_i["example"])
                final_prompt = self.task.construct_prompt(
                    input_i,
                    examples,
                    selected_labels=selected_labels,
                    max_input_tokens=self.llm.DEFAULT_CONTEXT_LENGTH,
                    get_num_tokens=self.llm.get_num_tokens,
                )
            else:
                final_prompt = self.task.construct_prompt(
                    input_i,
                    examples,
                    max_input_tokens=self.llm.DEFAULT_CONTEXT_LENGTH,
                    get_num_tokens=self.llm.get_num_tokens,
                )
            prompt_list.append(final_prompt)

            # Calculate the number of tokens
            curr_cost = self.llm.get_cost(prompt=final_prompt, label="")
            total_cost += curr_cost

        total_cost = total_cost * (len(dataset.inputs) / input_limit)
        table = {
            "Total Estimated Cost": f"${maybe_round(total_cost)}",
            "Number of Examples": len(dataset.inputs),
            "Average cost per example": f"${maybe_round(total_cost / len(dataset.inputs))}",
        }
        table = {"parameter": list(table.keys()), "value": list(table.values())}

        print_table(
            table, show_header=False, console=self.console, styles=COST_TABLE_STYLES
        )
        self.console.rule("Prompt Example")
        self.console.print(f"{prompt_list[0]}", markup=False)
        self.console.rule()

    async def async_run_transform(
        self, transform: BaseTransform, dataset: AutolabelDataset
    ):
        transform_outputs = [
            transform.apply(input_dict) for input_dict in dataset.inputs
        ]

        outputs = await gather_async_tasks_with_progress(
            transform_outputs,
            description=f"Running transform {transform.name()}...",
            console=self.console,
        )
        output_df = pd.DataFrame.from_records(outputs)
        final_df = pd.concat([dataset.df, output_df], axis=1)
        dataset = AutolabelDataset(final_df, self.config)
        return dataset

    def transform(self, dataset: AutolabelDataset):
        transforms = []
        for transform_dict in self.config.transforms():
            transforms.append(
                TransformFactory.from_dict(transform_dict, cache=self.transform_cache)
            )
        for transform in transforms:
            dataset = asyncio.run(self.async_run_transform(transform, dataset))

        return dataset

    async def get_confidence_score(
        self, annotation: LLMAnnotation, chunk: Dict, examples: List[Dict]
    ) -> Union[float, dict]:
        full_confidence_input = annotation.confidence_prompt + annotation.raw_response
        if (
            self.llm.returns_token_probs()
            or not self.config.confidence_chunk_column()
            or self.get_num_tokens(full_confidence_input)
            < self.CONFIDENCE_MAX_CONTEXT_LENGTH
        ):
            return await self.confidence.calculate(model_generation=annotation)
        key_to_chunk = self.config.confidence_chunk_column()
        if not key_to_chunk:
            raise ValueError(
                "confidence_chunk_column must be set in the config to use confidence_chunk_size"
            )
        if key_to_chunk == AUTO_CONFIDENCE_CHUNKING_COLUMN:
            # If the confidence_chunk_column is set to auto,
            # we choose the column with the most tokens as the chunking column.
            max_tokens = -1
            example_template_keys = get_format_variables(self.config.example_template())
            for key in example_template_keys:
                num_tokens = self.get_num_tokens(chunk[key])
                if num_tokens > max_tokens:
                    max_tokens = num_tokens
                    key_to_chunk = key

        empty_chunk = chunk.copy()
        empty_chunk[key_to_chunk] = ""
        empty_prompt = self.task.construct_confidence_prompt(empty_chunk, examples)
        num_tokens_empty_prompt = self.get_num_tokens(empty_prompt)
        num_tokens_per_chunk = (
            self.config.confidence_chunk_size() - num_tokens_empty_prompt
        )
        confidence_chunks = self.chunk_string(chunk[key_to_chunk], num_tokens_per_chunk)

        confidence_scores = []
        for confidence_chunk in confidence_chunks:
            new_chunk = chunk.copy()
            new_chunk[key_to_chunk] = confidence_chunk
            new_prompt = self.task.construct_confidence_prompt(new_chunk, examples)
            annotation_dict = annotation.dict()
            annotation_dict["confidence_prompt"] = new_prompt
            confidence_scores.append(
                await self.confidence.calculate(
                    model_generation=LLMAnnotation(**annotation_dict),
                )
            )

        merge_function = MERGE_FUNCTION[self.config.confidence_merge_function()]
        if isinstance(confidence_scores[0], dict):
            merged_confidence = {}
            for key in confidence_scores[0].keys():
                merged_confidence[key] = merge_function(
                    [conf[key] for conf in confidence_scores]
                )
            return merged_confidence
        else:
            merged_confidence = merge_function(confidence_scores)
            return merged_confidence

    def majority_annotation(
        self, annotation_list: List[LLMAnnotation]
    ) -> LLMAnnotation:
        labels = [a.label for a in annotation_list]
        counts = {}
        for ind, label in enumerate(labels):
            # Needed for named entity recognition which outputs lists instead of strings
            label = str(label)

            if label not in counts:
                counts[label] = (1, ind)
            else:
                counts[label] = (counts[label][0] + 1, counts[label][1])
        max_label = max(counts, key=lambda x: counts[x][0])
        return annotation_list[counts[max_label][1]]

    def generate_explanations(
        self,
        seed_examples: Union[str, List[Dict]],
        include_label: bool = True,
    ) -> List[Dict]:
        return asyncio.run(
            self.agenerate_explanations(
                seed_examples=seed_examples, include_label=include_label
            )
        )

    async def agenerate_explanations(
        self,
        seed_examples: Union[str, List[Dict]],
        include_label: bool = True,
    ) -> List[Dict]:
        """Use LLM to generate explanations for why examples are labeled the way that they are."""
        out_file = None
        if isinstance(seed_examples, str):
            out_file = seed_examples
            seed_loader = AutolabelDataset(seed_examples, self.config)
            seed_examples = seed_loader.inputs

        explanation_column = self.config.explanation_column()
        if not explanation_column:
            raise ValueError(
                "The explanation column needs to be specified in the dataset config."
            )

        for seed_example in track(
            seed_examples,
            description="Generating explanations",
            console=self.console,
        ):
            explanation_prompt = self.task.get_explanation_prompt(
                seed_example, include_label=include_label
            )
            if self.task.image_col is not None:
                explanation_prompt = json.dumps(
                    {
                        "text": explanation_prompt,
                        "image_url": seed_example[self.task.image_col],
                    }
                )
            explanation = await self.llm.label([explanation_prompt])
            explanation = explanation.generations[0][0].text
            seed_example[explanation_column] = str(explanation) if explanation else ""

        if out_file:
            df = pd.DataFrame.from_records(seed_examples)
            df.to_csv(out_file, index=False)

        return seed_examples

    def generate_synthetic_dataset(self) -> AutolabelDataset:
        columns = get_format_variables(self.config.example_template())
        df = pd.DataFrame(columns=columns)
        for label in track(
            self.config.labels_list(),
            description="Generating dataset",
            console=self.console,
        ):
            prompt = self.task.get_generate_dataset_prompt(label)

            result = self.llm.label([prompt])
            if result.errors[0] is not None:
                self.console.print(
                    f"Error generating rows for label {label}: {result.errors[0]}"
                )
            else:
                response = result.generations[0][0].text.strip()

                response = io.StringIO(response)
                label_df = pd.read_csv(response, sep=self.config.delimiter())
                label_df[self.config.label_column()] = label
                df = pd.concat([df, label_df], axis=0, ignore_index=True)
        return AutolabelDataset(df, self.config)

    def clear_cache(self, use_ttl: bool = True):
        """
        Clears the generation and transformation cache from autolabel.
        Args:
            use_ttl: If true, only clears the cache if the ttl has expired.
        """
        self.generation_cache.clear(use_ttl=use_ttl)
        self.transform_cache.clear(use_ttl=use_ttl)

    def get_num_tokens(self, inp: str) -> int:
        """Returns the number of tokens in the prompt"""
        return len(self.confidence_tokenizer.encode(str(inp)))

    def chunk_string(self, inp: str, chunk_size: int) -> List[str]:
        """Chunks the input string into chunks of size chunk_size"""
        tokens = self.confidence_tokenizer.encode(inp)
        chunks = [tokens[i : i + chunk_size] for i in range(0, len(tokens), chunk_size)]
        return [self.confidence_tokenizer.decode(chunk) for chunk in chunks]

agenerate_explanations(seed_examples, include_label=True) async

Use LLM to generate explanations for why examples are labeled the way that they are.

Source code in autolabel/src/autolabel/labeler.py
async def agenerate_explanations(
    self,
    seed_examples: Union[str, List[Dict]],
    include_label: bool = True,
) -> List[Dict]:
    """Use LLM to generate explanations for why examples are labeled the way that they are."""
    out_file = None
    if isinstance(seed_examples, str):
        out_file = seed_examples
        seed_loader = AutolabelDataset(seed_examples, self.config)
        seed_examples = seed_loader.inputs

    explanation_column = self.config.explanation_column()
    if not explanation_column:
        raise ValueError(
            "The explanation column needs to be specified in the dataset config."
        )

    for seed_example in track(
        seed_examples,
        description="Generating explanations",
        console=self.console,
    ):
        explanation_prompt = self.task.get_explanation_prompt(
            seed_example, include_label=include_label
        )
        if self.task.image_col is not None:
            explanation_prompt = json.dumps(
                {
                    "text": explanation_prompt,
                    "image_url": seed_example[self.task.image_col],
                }
            )
        explanation = await self.llm.label([explanation_prompt])
        explanation = explanation.generations[0][0].text
        seed_example[explanation_column] = str(explanation) if explanation else ""

    if out_file:
        df = pd.DataFrame.from_records(seed_examples)
        df.to_csv(out_file, index=False)

    return seed_examples

arun(dataset, output_name=None, max_items=None, start_index=0, additional_metrics=[], skip_eval=False) async

Labels data in a given dataset. Output written to new CSV file.

Parameters:

Name Type Description Default
dataset AutolabelDataset

path to CSV dataset to be annotated

required
max_items Optional[int]

maximum items in dataset to be annotated

None
output_name Optional[str]

custom name of output CSV file

None
start_index int

skips annotating [0, start_index)

0
Source code in autolabel/src/autolabel/labeler.py
async def arun(
    self,
    dataset: AutolabelDataset,
    output_name: Optional[str] = None,
    max_items: Optional[int] = None,
    start_index: int = 0,
    additional_metrics: Optional[List[BaseMetric]] = [],
    skip_eval: Optional[bool] = False,
) -> Tuple[pd.Series, pd.DataFrame, List[MetricResult]]:
    """Labels data in a given dataset. Output written to new CSV file.

    Args:
        dataset: path to CSV dataset to be annotated
        max_items: maximum items in dataset to be annotated
        output_name: custom name of output CSV file
        start_index: skips annotating [0, start_index)
    """

    dataset = dataset.get_slice(max_items=max_items, start_index=start_index)

    llm_labels = []

    # Get the seed examples from the dataset config
    seed_examples = self.config.few_shot_example_set()

    # If this dataset config is a string, read the corrresponding csv file
    if isinstance(seed_examples, str):
        seed_loader = AutolabelDataset(seed_examples, self.config)
        seed_examples = seed_loader.inputs

    # Check explanations are present in data if explanation_column is passed in
    if (
        self.config.explanation_column()
        and len(seed_examples) > 0
        and self.config.explanation_column() not in list(seed_examples[0].keys())
    ):
        raise ValueError(
            f"Explanation column {self.config.explanation_column()} not found in dataset.\nMake sure that explanations were generated using labeler.generate_explanations(seed_file)."
        )

    if self.example_selector is None:
        if (
            self.config.label_selection()
            and self.config.few_shot_algorithm() != "fixed"
        ):
            # TODO: Add support for other few shot algorithms specially semantic similarity
            raise ValueError(
                "Error: Only 'fixed' few shot example selector is supported for label selection."
            )

        self.example_selector = ExampleSelectorFactory.initialize_selector(
            self.config,
            [safe_serialize_to_string(example) for example in seed_examples],
            dataset.df.keys().tolist(),
            cache=self.generation_cache is not None,
        )

    if self.config.label_selection():
        if self.config.task_type() != TaskType.CLASSIFICATION:
            self.console.print(
                "Warning: label_selection only supported for classification tasks!"
            )
        else:
            self.label_selector = LabelSelector(
                config=self.config,
                embedding_func=PROVIDER_TO_MODEL.get(
                    self.config.embedding_provider(), DEFAULT_EMBEDDING_PROVIDER
                )(model=self.config.embedding_model_name()),
            )

    current_index = 0
    cost = 0.0
    postfix_dict = {}

    indices = range(current_index, len(dataset.inputs))
    selected_labels = self.config.labels_list()
    for current_index in track_with_stats(
        indices,
        postfix_dict,
        total=len(dataset.inputs) - current_index,
        console=self.console,
    ):
        chunk = dataset.inputs[current_index]
        examples = []

        if (
            self.config.label_selection()
            and self.config.task_type() == TaskType.CLASSIFICATION
        ):
            # get every column except the one we want to label
            toEmbed = chunk.copy()
            if self.config.label_column() and self.config.label_column() in toEmbed:
                del toEmbed[self.config.label_column()]

            # convert this to a string
            toEmbed = json.dumps(toEmbed)

            selected_labels = self.label_selector.select_labels(toEmbed)

            if self.example_selector:
                examples = self.example_selector.select_examples(
                    safe_serialize_to_string(chunk),
                    selected_labels=selected_labels,
                    label_column=self.config.label_column(),
                )
            else:
                examples = []
        else:
            if self.example_selector:
                examples = self.example_selector.select_examples(
                    safe_serialize_to_string(chunk),
                )

        # Construct Prompt to pass to LLM
        final_prompt = self.task.construct_prompt(
            chunk,
            examples,
            selected_labels=selected_labels,
            max_input_tokens=self.llm.DEFAULT_CONTEXT_LENGTH,
            get_num_tokens=self.llm.get_num_tokens,
        )

        response = await self.llm.label([final_prompt])
        for i, generations, error, latency in zip(
            range(len(response.generations)),
            response.generations,
            response.errors,
            response.latencies,
        ):
            input_tokens = self.llm.get_num_tokens(final_prompt)
            if error is not None:
                annotation = LLMAnnotation(
                    successfully_labeled=False,
                    label=self.task.NULL_LABEL_TOKEN,
                    raw_response="",
                    curr_sample=pickle.dumps(chunk),
                    prompt=final_prompt,
                    confidence_score=0,
                    error=error,
                    input_tokens=input_tokens,
                    cost=0,
                    latency=0,
                )
            else:
                annotations = []
                for generation in generations:
                    annotation = self.task.parse_llm_response(
                        generation, chunk, final_prompt
                    )
                    annotation.confidence_prompt = (
                        self.task.construct_confidence_prompt(chunk, examples)
                    )
                    annotation.input_tokens = input_tokens
                    annotation.output_tokens = self.llm.get_num_tokens(
                        annotation.raw_response
                    )
                    annotation.cost = sum(response.costs)
                    annotation.latency = latency

                    if self.config.confidence():
                        try:
                            annotation.confidence_score = (
                                await self.get_confidence_score(
                                    annotation, chunk, examples
                                )
                            )
                        except Exception as e:
                            logger.exception(
                                f"Error calculating confidence score: {e}"
                            )
                            logger.warning(
                                f"Could not calculate confidence score for annotation: {annotation}"
                            )
                            if (
                                self.config.task_type()
                                == TaskType.ATTRIBUTE_EXTRACTION
                            ):
                                annotation.confidence_score = {}
                            else:
                                annotation.confidence_score = 0

                    annotations.append(annotation)
                annotation = self.majority_annotation(annotations)

            llm_labels.append(annotation)

        cost += sum(response.costs)
        postfix_dict[self.COST_KEY] = f"{cost:.2f}"

        # Evaluate the task every eval_every examples
        if not skip_eval and (current_index + 1) % 100 == 0:
            if dataset.gt_labels:
                eval_result = self.task.eval(
                    llm_labels,
                    (
                        dataset.gt_labels[: len(llm_labels)]
                        if isinstance(dataset.gt_labels, list)
                        else {
                            k: v[: len(llm_labels)]
                            for k, v in dataset.gt_labels.items()
                        }
                    ),
                    additional_metrics=additional_metrics,
                )

                for m in eval_result:
                    # This is a row wise metric
                    if isinstance(m.value, list):
                        continue
                    elif m.show_running:
                        postfix_dict[m.name] = (
                            f"{m.value:.4f}"
                            if isinstance(m.value, float)
                            else m.value
                        )

    eval_result = None
    table = {}

    # if true labels are provided, evaluate accuracy of predictions
    if not skip_eval and dataset.gt_labels:
        eval_result = self.task.eval(
            llm_labels,
            (
                dataset.gt_labels[: len(llm_labels)]
                if isinstance(dataset.gt_labels, list)
                else {k: v[: len(llm_labels)] for k, v in dataset.gt_labels.items()}
            ),
            additional_metrics=additional_metrics,
        )
        # TODO: serialize and write to file
        for m in eval_result:
            if isinstance(m.value, list):
                continue
            elif m.show_running:
                table[m.name] = m.value
            else:
                self.console.print(f"{m.name}:\n{m.value}")

    # print cost
    self.console.print(f"Actual Cost: {maybe_round(cost)}")
    print_table(table, console=self.console, default_style=METRIC_TABLE_STYLE)

    dataset.process_labels(llm_labels, eval_result)
    # Only save to csv if output_name is provided or dataset is a string
    if not output_name and isinstance(dataset, str):
        output_name = (
            dataset.rsplit(".", 1)[0] + "_labeled." + dataset.rsplit(".", 1)[1]
        )

    if output_name:
        dataset.save(output_file_name=output_name)
    return dataset

chunk_string(inp, chunk_size)

Chunks the input string into chunks of size chunk_size

Source code in autolabel/src/autolabel/labeler.py
def chunk_string(self, inp: str, chunk_size: int) -> List[str]:
    """Chunks the input string into chunks of size chunk_size"""
    tokens = self.confidence_tokenizer.encode(inp)
    chunks = [tokens[i : i + chunk_size] for i in range(0, len(tokens), chunk_size)]
    return [self.confidence_tokenizer.decode(chunk) for chunk in chunks]

clear_cache(use_ttl=True)

Clears the generation and transformation cache from autolabel. Args: use_ttl: If true, only clears the cache if the ttl has expired.

Source code in autolabel/src/autolabel/labeler.py
def clear_cache(self, use_ttl: bool = True):
    """
    Clears the generation and transformation cache from autolabel.
    Args:
        use_ttl: If true, only clears the cache if the ttl has expired.
    """
    self.generation_cache.clear(use_ttl=use_ttl)
    self.transform_cache.clear(use_ttl=use_ttl)

get_num_tokens(inp)

Returns the number of tokens in the prompt

Source code in autolabel/src/autolabel/labeler.py
def get_num_tokens(self, inp: str) -> int:
    """Returns the number of tokens in the prompt"""
    return len(self.confidence_tokenizer.encode(str(inp)))

plan(dataset, max_items=None, start_index=0)

Calculates and prints the cost of calling autolabel.run() on a given dataset

Parameters:

Name Type Description Default
dataset AutolabelDataset

path to a CSV dataset

required
Source code in autolabel/src/autolabel/labeler.py
def plan(
    self,
    dataset: AutolabelDataset,
    max_items: Optional[int] = None,
    start_index: int = 0,
) -> None:
    """Calculates and prints the cost of calling autolabel.run() on a given dataset

    Args:
        dataset: path to a CSV dataset
    """
    dataset = dataset.get_slice(max_items=max_items, start_index=start_index)

    if (
        self.config.confidence()
        and "REFUEL_API_KEY" not in os.environ
        and not self.llm.returns_token_probs()
    ):
        raise ValueError(
            "REFUEL_API_KEY environment variable must be set to compute confidence scores. You can request an API key at https://refuel-ai.typeform.com/llm-access."
        )

    prompt_list = []
    total_cost = 0

    # Get the seed examples from the dataset config
    seed_examples = self.config.few_shot_example_set()

    # If this dataset config is a string, read the corrresponding csv file
    if isinstance(seed_examples, str):
        seed_loader = AutolabelDataset(seed_examples, self.config)
        seed_examples = seed_loader.inputs

    # Check explanations are present in data if explanation_column is passed in
    if (
        self.config.explanation_column()
        and len(seed_examples) > 0
        and self.config.explanation_column() not in list(seed_examples[0].keys())
    ):
        raise ValueError(
            f"Explanation column {self.config.explanation_column()} not found in dataset.\nMake sure that explanations were generated using labeler.generate_explanations(seed_file)."
        )

    self.example_selector = ExampleSelectorFactory.initialize_selector(
        self.config,
        [safe_serialize_to_string(example) for example in seed_examples],
        dataset.df.keys().tolist(),
        cache=self.generation_cache is not None,
    )

    if self.config.label_selection():
        if self.config.task_type() != TaskType.CLASSIFICATION:
            self.console.print(
                "Warning: label_selection only supported for classification tasks!"
            )
        else:
            self.label_selector = LabelSelector(
                config=self.config,
                embedding_func=PROVIDER_TO_MODEL.get(
                    self.config.embedding_provider(), DEFAULT_EMBEDDING_PROVIDER
                )(model=self.config.embedding_model_name()),
            )

    input_limit = min(len(dataset.inputs), 100) if max_items is None else max_items  # type: ignore
    for input_i in track(
        dataset.inputs[:input_limit],
        description="Generating Prompts...",
        console=self.console,
    ):
        # TODO: Check if this needs to use the example selector
        if self.example_selector:
            examples = self.example_selector.select_examples(
                safe_serialize_to_string(input_i)
            )
        else:
            examples = []
        if (
            self.config.label_selection()
            and self.config.task_type() == TaskType.CLASSIFICATION
        ):
            selected_labels = self.label_selector.select_labels(input_i["example"])
            final_prompt = self.task.construct_prompt(
                input_i,
                examples,
                selected_labels=selected_labels,
                max_input_tokens=self.llm.DEFAULT_CONTEXT_LENGTH,
                get_num_tokens=self.llm.get_num_tokens,
            )
        else:
            final_prompt = self.task.construct_prompt(
                input_i,
                examples,
                max_input_tokens=self.llm.DEFAULT_CONTEXT_LENGTH,
                get_num_tokens=self.llm.get_num_tokens,
            )
        prompt_list.append(final_prompt)

        # Calculate the number of tokens
        curr_cost = self.llm.get_cost(prompt=final_prompt, label="")
        total_cost += curr_cost

    total_cost = total_cost * (len(dataset.inputs) / input_limit)
    table = {
        "Total Estimated Cost": f"${maybe_round(total_cost)}",
        "Number of Examples": len(dataset.inputs),
        "Average cost per example": f"${maybe_round(total_cost / len(dataset.inputs))}",
    }
    table = {"parameter": list(table.keys()), "value": list(table.values())}

    print_table(
        table, show_header=False, console=self.console, styles=COST_TABLE_STYLES
    )
    self.console.rule("Prompt Example")
    self.console.print(f"{prompt_list[0]}", markup=False)
    self.console.rule()