[Benchmark Backfill] Integrate OfficeQA into lmms-eval

docs/current_tasks.md

@@ -402,6 +402,7 @@ python -m lmms_eval --tasks list_with_num
 - [Infographic VQA](https://www.docvqa.org/datasets/infographicvqa) (infovqa)
   - Infographic VQA Validation (infovqa_val)
   - Infographic VQA Test (infovqa_test)
+- [OfficeQA](https://github.com/databricks/officeqa) (officeqa)
 - [OCRBench](https://github.com/Yuliang-Liu/MultimodalOCR) (ocrbench)
 - [OCRBench v2](https://github.com/Yuliang-Liu/MultimodalOCR) (ocrbench_v2)
 - [PRISMM-Bench](https://arxiv.org/abs/2510.16505) (prismm_bench)

lmms_eval/tasks/officeqa/officeqa.yaml

@@ -0,0 +1,27 @@
+dataset_path: csv
+dataset_kwargs:
+  data_files: https://raw.githubusercontent.com/databricks/officeqa/main/officeqa.csv
+task: officeqa
+test_split: train
+output_type: generate_until
+doc_to_visual: !function utils.officeqa_doc_to_visual
+doc_to_text: !function utils.officeqa_doc_to_text
+doc_to_target: !function utils.officeqa_doc_to_target
+generation_kwargs:
+  max_new_tokens: 64
+  temperature: 0
+  do_sample: false
+process_results: !function utils.officeqa_process_results
+metric_list:
+  - metric: officeqa_relaxed_accuracy
+    aggregation: mean
+    higher_is_better: true
+lmms_eval_specific_kwargs:
+  default:
+    pre_prompt: ""
+    post_prompt: "\nAnswer with a single word, phrase, or option letter."
+  qwen_vl:
+    pre_prompt: ""
+    post_prompt: " Answer:"
+metadata:
+  - version: 0.0

lmms_eval/tasks/officeqa/utils.py

@@ -0,0 +1,156 @@
+import io
+import re
+from typing import Any
+
+from PIL import Image
+
+
+def _normalize_text(text: Any) -> str:
+    if text is None:
+        return ""
+    return " ".join(str(text).strip().lower().split())
+
+
+def _to_float(text: str):
+    candidate = text.strip().replace(",", "")
+    if candidate.endswith("%"):
+        try:
+            return float(candidate[:-1]) / 100.0
+        except ValueError:
+            return None
+
+    match = re.match(r"^[-+]?\d*\.?\d+$", candidate)
+    if not match:
+        return None
+
+    try:
+        return float(candidate)
+    except ValueError:
+        return None
+
+
+def _relaxed_match(prediction: str, target: str, max_relative_change: float = 0.05) -> float:
+    pred = _normalize_text(prediction)
+    tgt = _normalize_text(target)
+    if not pred or not tgt:
+        return 0.0
+
+    pred_float = _to_float(pred)
+    tgt_float = _to_float(tgt)
+
+    if pred_float is not None and tgt_float is not None:
+        if tgt_float == 0:
+            return float(pred_float == 0)
+        return float(abs(pred_float - tgt_float) / abs(tgt_float) <= max_relative_change)
+
+    return float(pred == tgt)
+
+
+def _to_rgb(image_obj: Any):
+    if isinstance(image_obj, Image.Image):
+        return image_obj.convert("RGB")
+
+    if isinstance(image_obj, bytes):
+        return Image.open(io.BytesIO(image_obj)).convert("RGB")
+
+    return None
+
+
+def _extract_question(doc: dict) -> str:
+    for key in ["question", "query", "prompt", "instruction"]:
+        value = doc.get(key)
+        if isinstance(value, str) and value.strip():
+            return value.strip()
+    return ""
+
+
+def _extract_options(doc: dict) -> list[str]:
+    options = doc.get("options", doc.get("choices"))
+    if isinstance(options, list):
+        normalized = []
+        for item in options:
+            if isinstance(item, dict):
+                text = item.get("text", item.get("option", ""))
+                normalized.append(str(text))
+            else:
+                normalized.append(str(item))
+        return [x for x in normalized if x.strip()]
+    return []
+
+
+def _extract_answers(doc: dict) -> list[str]:
+    answers = doc.get("answers", doc.get("answer", doc.get("target")))
+    if answers is None:
+        return []
+    if isinstance(answers, list):
+        return [str(item) for item in answers if str(item).strip()]
+    return [str(answers)]
+
+
+def _extract_option_letter(prediction: str) -> str:
+    normalized = prediction.strip().upper()
+    match = re.search(r"\b([A-Z])\b", normalized)
+    if match:
+        return match.group(1)
+    return ""
+
+
+def officeqa_doc_to_visual(doc):
+    visuals = []
+
+    for key in ["image", "page_image", "document_image"]:
+        if key in doc:
+            img = _to_rgb(doc[key])
+            if img is not None:
+                visuals.append(img)
+
+    for key in ["images", "page_images", "document_images", "pages"]:
+        value = doc.get(key)
+        if isinstance(value, list):
+            for item in value:
+                img = _to_rgb(item)
+                if img is not None:
+                    visuals.append(img)
+
+    return visuals
+
+
+def officeqa_doc_to_text(doc, lmms_eval_specific_kwargs=None):
+    kwargs = lmms_eval_specific_kwargs or {}
+    pre_prompt = kwargs.get("pre_prompt", "")
+    post_prompt = kwargs.get("post_prompt", "")
+
+    question = _extract_question(doc)
+    options = _extract_options(doc)
+    if not options:
+        return f"{pre_prompt}{question}{post_prompt}"
+
+    option_labels = [chr(ord("A") + idx) for idx in range(len(options))]
+    option_lines = "\n".join(f"{label}. {choice}" for label, choice in zip(option_labels, options))
+    return f"{pre_prompt}{question}\n{option_lines}{post_prompt}"
+
+
+def officeqa_doc_to_target(doc):
+    answers = _extract_answers(doc)
+    return answers[0] if answers else ""
+
+
+def officeqa_process_results(doc, results):
+    prediction = str(results[0]).strip()
+    answers = _extract_answers(doc)
+
+    if not answers:
+        return {"officeqa_relaxed_accuracy": 0.0}
+
+    best_score = max(_relaxed_match(prediction, answer) for answer in answers)
+
+    options = _extract_options(doc)
+    if options:
+        pred_letter = _extract_option_letter(prediction)
+        if pred_letter:
+            for answer in answers:
+                answer_norm = answer.strip().upper()
+                if pred_letter == answer_norm[:1]:
+                    best_score = max(best_score, 1.0)
+
+    return {"officeqa_relaxed_accuracy": best_score}