Merge pull request rllm-org#106 from evanmiller-anthropic/docvqa

DocVQA implementation

Author: MattFisher

README.md

@@ -300,6 +300,14 @@ The questions were generated by GPT-4 based on the "Computer Systems Security: P
    inspect eval inspect_evals/boolq
    ```
 
+- ### [DocVQA: A Dataset for VQA on Document Images](src/inspect_evals/docvqa)
+  DocVQA is a Visual Question Answering benchmark that consists of 50,000 questions covering 12,000+ document images. This implementation solves and scores the "validation" split.
+ <sub><sup>Contributed by: [@evanmiller-anthropic](https://github.com/evanmiller-anthropic)</sub></sup>
+
+   ```bash
+   inspect eval inspect_evals/docvqa
+   ```
+
 - ### [DROP: A Reading Comprehension Benchmark Requiring Discrete Reasoning Over Paragraphs](src/inspect_evals/drop)
   Evaluates reading comprehension where models must resolve references in a question, perhaps to multiple input positions, and perform discrete operations over them (such as addition, counting, or sorting).
  <sub><sup>Contributed by: [@xeon27](https://github.com/xeon27)</sub></sup>
@@ -443,4 +451,4 @@ The questions were generated by GPT-4 based on the "Computer Systems Security: P
    inspect eval inspect_evals/agie_lsat_lr
    ```
 
-<!-- /Eval Listing: Automatically Generated -->
+<!-- /Eval Listing: Automatically Generated -->

src/inspect_evals/_registry.py

@@ -26,6 +26,7 @@
     cybermetric_10000,
 )
 from .cyberseceval_2 import interpreter_abuse, prompt_injection, vulnerability_exploit
+from .docvqa import docvqa
 from .drop import drop
 from .ds1000 import ds1000
 from .gaia import gaia, gaia_level1, gaia_level2, gaia_level3

src/inspect_evals/docvqa/README.md

@@ -0,0 +1,73 @@
+# DocVQA: A Dataset for VQA on Document Images
+
+[DocVQA](https://arxiv.org/abs/2007.00398) is a Visual Question Answering benchmark that consists of 50,000 questions covering 12,000+ document images. This implementation solves and scores the "validation" split.
+
+<!-- Contributors: Automatically Generated -->
+Contributed by [@evanmiller-anthropic](https://github.com/evanmiller-anthropic)
+<!-- /Contributors: Automatically Generated -->
+
+
+<!-- Usage: Automatically Generated -->
+## Usage
+
+First, install the `inspect_ai` and `inspect_evals` Python packages with:
+
+```bash
+pip install inspect_ai
+pip install git+https://github.com/UKGovernmentBEIS/inspect_evals
+```
+
+Then, evaluate against one or more models with:
+
+```bash
+inspect eval inspect_evals/docvqa --model openai/gpt-4o
+```
+
+After running evaluations, you can view their logs using the `inspect view` command:
+
+```bash
+inspect view
+```
+
+If you don't want to specify the `--model` each time you run an evaluation, create a `.env` configuration file in your working directory that defines the `INSPECT_EVAL_MODEL` environment variable along with your API key. For example:
+
+```bash
+INSPECT_EVAL_MODEL=anthropic/claude-3-5-sonnet-20240620
+ANTHROPIC_API_KEY=<anthropic-api-key>
+```
+<!-- /Usage: Automatically Generated -->
+
+<!-- Options: Automatically Generated -->
+## Options
+
+You can control a variety of options from the command line. For example:
+
+```bash
+inspect eval inspect_evals/docvqa --limit 10
+inspect eval inspect_evals/docvqa --max-connections 10
+inspect eval inspect_evals/docvqa --temperature 0.5
+```
+
+See `inspect eval --help` for all available options.
+<!-- /Options: Automatically Generated -->
+
+## Dataset
+
+The DocVQA dataset contains a "validation" split and a "test" split. To prevent leakage into training data, the authors of DocVQA have chosen to hold back the answers to the "test" split. Scoring on the "test" split requires coordinating with the DocVQA authors.
+
+Each split contains several questions about each image. Here is an example image:
+
+![Diabetes in Australia](https://rrc.cvc.uab.es/files/DocVQA_exT3_2_Infographics.png)
+
+And associated example questions:
+* How many females are affected by diabetes?
+* What percentage of cases can be prevented?
+* What could lead to blindness or stroke diabetes?
+
+The model is tasked to answer each question by referring to the image. The prompts are based on OpenAI's [simple-evals](https://github.com/openai/simple-evals/blob/294cb1f/drop_eval.py#L261C13-L283C91).
+
+## Scoring
+
+DocVQA computes the Average Normalized Levenstein Similarity:
+
+[Average Normalized Levenstein Similarity definition](https://user-images.githubusercontent.com/48327001/195277520-b1ef2be2-c4d7-417b-91ec-5fda8aa6db06.png)

src/inspect_evals/docvqa/__init__.py

@@ -0,0 +1,3 @@
+from .docvqa import docvqa
+
+__all__ = ["docvqa"]

src/inspect_evals/docvqa/docvqa.py

@@ -0,0 +1,188 @@
+import re
+from io import BytesIO
+from pathlib import Path
+from typing import Any
+
+from inspect_ai import Task, task
+from inspect_ai.dataset import Sample, hf_dataset
+from inspect_ai.model import ChatMessage, ChatMessageUser, ContentImage, ContentText
+from inspect_ai.scorer import (
+    INCORRECT,
+    AnswerPattern,
+    Score,
+    Scorer,
+    Target,
+    accuracy,
+    scorer,
+    stderr,
+)
+from inspect_ai.solver import (
+    Generate,
+    Solver,
+    TaskState,
+    solver,
+)
+from PIL import Image
+from platformdirs import user_cache_dir
+
+FREEFORM_TEMPLATE = r"""
+Answer the following question. The entire content of your response should be of the following format: 'ANSWER: $ANSWER' (without quotes) where $ANSWER is your answer.
+
+{question}
+"""
+
+IMAGE_BASE_DIR = Path(user_cache_dir("inspect_evals")) / "docvqa_images"
+
+
+def _levenshtein_distance(str1: str, str2: str) -> int:
+    """Computes a Levenshtein distance, same as Levenshtein.distance in the python-Levenshtein package."""
+    # Create a matrix of size (len(str1) + 1) x (len(str2) + 1)
+    matrix = [[0 for j in range(len(str2) + 1)] for i in range(len(str1) + 1)]
+
+    # Initialize the first row and column
+    for i in range(len(str1) + 1):
+        matrix[i][0] = i
+    for j in range(len(str2) + 1):
+        matrix[0][j] = j
+
+    # Fill in the rest of the matrix
+    for i in range(1, len(str1) + 1):
+        for j in range(1, len(str2) + 1):
+            matrix[i][j] = min(
+                matrix[i - 1][j] + 1,  # deletion
+                matrix[i][j - 1] + 1,  # insertion
+                matrix[i - 1][j - 1] + int(str1[i - 1] != str2[j - 1]),  # substitution
+            )
+
+    return matrix[len(str1)][len(str2)]
+
+
+def _best_normalized_levenshtein_similiarity(
+    completion: str, ground_truths: list[str], threshold: float
+) -> float:
+    """
+    Compute a best normalized Levenshtein similiarity, an input into the Average Normalized Levenshtein Similiarity (ANLS)
+
+    The Average Normalized Levenshtein Similarity (ANLS) is defined in equation (1) of
+    https://arxiv.org/pdf/1907.00490.pdf
+
+    Note that the "average" is computed by the accuracy metric -- not here. This function computes
+    the term inside the summation of equation (1).
+    """
+    best_score = 0.0
+    for ground_truth in ground_truths:
+        if len(ground_truth) == 0 and len(completion) == 0:
+            best_score = 1
+            break
+        levenshtein_distance = _levenshtein_distance(
+            completion.lower(), ground_truth.lower()
+        )
+        normed_levenshtein_distance = levenshtein_distance / max(
+            len(completion), len(ground_truth)
+        )
+        if normed_levenshtein_distance < threshold:
+            score = 1.0 - normed_levenshtein_distance
+        else:
+            score = 0.0
+        if score > best_score:
+            best_score = score
+    return best_score
+
+
+@task
+def docvqa() -> Task:
+    dataset = hf_dataset(
+        path="lmms-lab/DocVQA",
+        name="DocVQA",
+        split="validation",  # "answers" in the "test" split are held back by the authors
+        sample_fields=record_to_sample,
+        trust=True,
+        shuffle=True,
+    )
+
+    return Task(
+        dataset=dataset,
+        solver=[docvqa_solver()],
+        scorer=docvqa_scorer(),
+    )
+
+
+@scorer(metrics=[accuracy(), stderr()])
+def docvqa_scorer() -> Scorer:
+    async def normalized_levenshtein_similiarity_score(
+        state: TaskState, target: Target
+    ) -> Score:
+        threshold = 0.5
+        ground_truths = target.target
+        match = re.search(
+            AnswerPattern.LINE,
+            state.output.completion,
+            re.IGNORECASE,
+        )
+        if match:
+            completion = match.groups()[0]
+            return Score(
+                value=_best_normalized_levenshtein_similiarity(
+                    completion, ground_truths, threshold
+                ),
+                answer=completion,
+            )
+
+        else:
+            # didn't find the scoring pattern
+            return Score(
+                value=INCORRECT,
+                explanation="Scoring pattern not matched in output: "
+                + f"{state.output.completion}",
+            )
+
+    return normalized_levenshtein_similiarity_score
+
+
+@solver
+def docvqa_solver() -> Solver:
+    async def solve(state: TaskState, generate: Generate) -> TaskState:
+        state.user_prompt.text = FREEFORM_TEMPLATE.format(
+            question=state.user_prompt.text
+        )
+        return await generate(state)
+
+    return solve
+
+
+def record_to_sample(record: dict[str, Any]) -> Sample:
+    # extract image
+    image_path = Path(IMAGE_BASE_DIR / record["image"]["path"])
+
+    image_bytes = record["image"]["bytes"]
+    assert is_image_png(image_bytes)
+
+    if not image_path.exists():
+        print(f"Extracting {image_path.name}")
+        # ensure parent
+        image_path.parent.mkdir(exist_ok=True, parents=True)
+        # reduce the image size
+        img = Image.open(BytesIO(image_bytes))
+        img.thumbnail((1024, 1024))
+        # save preserving format
+        img.save(image_path, format=img.format)
+
+    message: list[ChatMessage] = [
+        ChatMessageUser(
+            content=[
+                ContentText(text=record["question"]),
+                ContentImage(image=image_path.as_posix()),
+            ]
+        )
+    ]
+
+    return Sample(
+        input=message,
+        target=record["answers"],
+        id=record["questionId"],
+        metadata={"document_id": record["docId"]},
+    )
+
+
+def is_image_png(image_bytes: bytes) -> bool:
+    return image_bytes[:8] == b"\x89\x50\x4e\x47\x0d\x0a\x1a\x0a"

tests/docvqa/test_docvqa.py

@@ -0,0 +1,97 @@
+from inspect_evals.docvqa.docvqa import (
+    _best_normalized_levenshtein_similiarity as best,
+)
+from inspect_evals.docvqa.docvqa import (
+    _levenshtein_distance as levenshtein,
+)
+
+
+def test_levenshtein():
+    # Basic test cases
+    assert levenshtein("", "") == 0  # Empty strings
+    assert levenshtein("a", "a") == 0  # Same single char
+    assert levenshtein("abc", "abc") == 0  # Same string
+
+    # Single operations
+    assert levenshtein("a", "") == 1  # Single deletion
+    assert levenshtein("", "a") == 1  # Single insertion
+    assert levenshtein("a", "b") == 1  # Single substitution
+
+    # Multiple operations
+    assert levenshtein("kitten", "sitting") == 3  # Classic example
+    assert levenshtein("sunday", "saturday") == 3  # Real words
+
+
+def test_best_normalized_levenshtein_distance():
+    def best_norm_lev_sim(completion, ground_truths, threshold=2.0):
+        return round(best(completion, ground_truths, threshold), 3)
+
+    # Basic cases
+    assert best_norm_lev_sim("", [""]) == 1.0  # Empty strings
+    assert best_norm_lev_sim("a", ["a"]) == 1.0  # Single char match
+    assert best_norm_lev_sim("", ["a"]) == 0.0  # Empty vs char
+    assert best_norm_lev_sim("a", ["b"]) == 0.0  # Different chars
+
+    # Multiple correct answers
+    assert (
+        best_norm_lev_sim("color", ["color", "colour"]) == 1.0
+    )  # Exact match with variants
+
+    assert (
+        best_norm_lev_sim("theatre", ["theater", "theatre"]) == 1.0
+    )  # Regional spellings
+
+    # Partial matches with multiple answers
+    assert best_norm_lev_sim("thetre", ["theater", "theatre"]) == round(
+        1 - 1 / 7, 3
+    )  # One deletion
+
+    # Case insensitivity
+    assert best_norm_lev_sim("HELLO", ["hello", "hola"]) == 1.0  # All case differences
+
+    # Length differences
+    assert best_norm_lev_sim("hi", ["hello", "hey"]) == round(
+        1 - 2 / 3, 3
+    )  # Short vs longer options
+
+    assert best_norm_lev_sim("hi", ["hello", "hey"], 0.5) == 0.0  # Test threshold
+
+    assert best_norm_lev_sim("hi", ["hello", "hey"], 0.75) == round(
+        1 - 2 / 3, 3
+    )  # Test threshold
+
+    # Numeric and special characters
+    assert (
+        best_norm_lev_sim("2nd floor", ["second floor", "2nd floor", "floor 2"]) == 1.0
+    )  # Number representations
+
+    # Common abbreviations
+    assert (
+        best_norm_lev_sim("dept", ["department", "dept.", "dept"]) == 1.0
+    )  # Abbreviation matches
+
+    # Multiple errors
+    assert best_norm_lev_sim(
+        "californa", ["california", "calif", "ca"]
+    ) > best_norm_lev_sim(
+        "calfrnia", ["california", "calif", "ca"]
+    )  # Better partial match
+
+    # Spaces and formatting
+    assert (
+        best_norm_lev_sim("new york", ["newyork", "new york", "ny"]) == 1.0
+    )  # Space variations
+
+    # Unicode and special characters
+    assert best_norm_lev_sim("café", ["cafe", "café", "caffè"]) == 1.0  # Accent marks
+
+    # Long string comparisons
+    assert (
+        best_norm_lev_sim(
+            "mississipi river", ["mississippi river", "river mississippi"]
+        )
+        > 0.9
+    )  # Minor spelling error
+
+    # Completely different strings
+    assert best_norm_lev_sim("kiwi", ["banana", "orange"]) == 0.0  # No similarity