feat: Add AI2D task and refactor registries to use decorator pattern

- Add new AI2D task that loads the lmms-lab/ai2d dataset
- Refactor task registry to use @register_task decorator pattern
- Refactor scorer registry to use @register_scorer decorator pattern
- Update all task and scorer classes to use the new decorators
- Support multiple registration names for each task/scorer
- Maintain full backward compatibility with existing API

This change eliminates duplication between __init__.py and registry files,
making it easier to add new tasks and scorers.

🤖 Generated with [Claude Code](https://claude.ai/code)

Co-Authored-By: Claude <[email protected]>

Author: silviase

src/eval_mm/metrics/cc_ocr_scorer.py

@@ -3,6 +3,7 @@
 from typing import List, Dict, Any, cast  # Added cast for type hinting clarity
 
 from .scorer import Scorer, AggregateOutput, ScorerConfig
+from .scorer_registry import register_scorer
 
 
 def token_normalize(
@@ -158,6 +159,7 @@ def calculate_metrics(
 
 
 # CCOCRScorer class, specialized for Japanese (character-level, no alphanum_only)
+@register_scorer("cc-ocr", "CC-OCR", "CCOCRScorer")
 class CCOCRScorer(Scorer):
     def __init__(self, config: ScorerConfig):
         super().__init__(config)

src/eval_mm/metrics/exact_match_scorer.py

@@ -1,6 +1,8 @@
 from .scorer import Scorer, AggregateOutput
+from .scorer_registry import register_scorer
 
 
+@register_scorer("exact-match", "ExactMatch", "ExactMatchScorer")
 class ExactMatchScorer(Scorer):
     @staticmethod
     def score(refs: list[str], preds: list[str]) -> list[int]:

src/eval_mm/metrics/heron_bench_scorer.py

@@ -2,6 +2,7 @@
 from collections import defaultdict
 import numpy as np
 from eval_mm.metrics.scorer import Scorer, AggregateOutput
+from .scorer_registry import register_scorer
 import re
 import json
 
@@ -110,6 +111,7 @@ def ask_gpt4_batch(
     return completions
 
 
+@register_scorer("heron-bench", "HeronBench", "HeronBenchScorer")
 class HeronBenchScorer(Scorer):
     def score(self, refs, preds: list[str]) -> list[dict[str, int]]:
         docs = self.config.docs

src/eval_mm/metrics/jdocqa_scorer.py

@@ -1,4 +1,5 @@
 from eval_mm.metrics.scorer import Scorer, AggregateOutput
+from .scorer_registry import register_scorer
 from sacrebleu import sentence_bleu
 from unicodedata import normalize
 
@@ -38,6 +39,7 @@ def bleu_ja(refs, pred):
     return bleu_score.score / 100
 
 
+@register_scorer("jdocqa", "JDocQA", "JDocQAScorer")
 class JDocQAScorer(Scorer):
     def score(self, refs: list[str], preds: list[str]) -> list[int]:
         docs = self.config.docs

src/eval_mm/metrics/jic_vqa_scorer.py

@@ -1,6 +1,8 @@
 from .scorer import Scorer, AggregateOutput
+from .scorer_registry import register_scorer
 
 
+@register_scorer("jic-vqa", "JICVQA", "JICVQAScorer")
 class JICVQAScorer(Scorer):
     @staticmethod
     def score(refs: list[str], preds: list[str]) -> list[int]:

src/eval_mm/metrics/jmmmu_scorer.py

@@ -4,6 +4,7 @@
 import numpy as np
 from datasets import Dataset
 from eval_mm.metrics.scorer import Scorer, AggregateOutput
+from .scorer_registry import register_scorer
 
 
 DOMAIN_CAT2SUB_CAT = {
@@ -411,6 +412,7 @@ def get_score(doc: Dataset, pred: str, random_choice: bool = False) -> int:
     return score
 
 
+@register_scorer("jmmmu", "JMMMU", "JMMMUScorer")
 class JMMMUScorer(Scorer):
     def score(self, refs: list[str], preds: list[str]) -> list[int]:
         docs = self.config.docs

src/eval_mm/metrics/llm_as_a_judge_scorer.py

@@ -1,4 +1,5 @@
 from eval_mm.metrics.scorer import Scorer, AggregateOutput
+from .scorer_registry import register_scorer
 from tqdm import tqdm
 import re
 
@@ -24,6 +25,7 @@
 """
 
 
+@register_scorer("llm-as-a-judge", "LLM-as-a-Judge", "LlmAsaJudgeScorer")
 class LlmAsaJudgeScorer(Scorer):
     def score(
         self,

src/eval_mm/metrics/mecha_ja_scorer.py

@@ -1,5 +1,6 @@
 # mecha-ja-scorer.py
 from .scorer import Scorer, AggregateOutput
+from .scorer_registry import register_scorer
 import re
 from collections import defaultdict
 
@@ -9,6 +10,7 @@
 }
 
 
+@register_scorer("mecha-ja", "MECHAJa", "MECHAJaScorer")
 class MECHAJaScorer(Scorer):
     @staticmethod
     def _parse_rotation_id(qid: str) -> str:

src/eval_mm/metrics/mmmu_scorer.py

@@ -4,6 +4,7 @@
 import numpy as np
 from datasets import Dataset
 from .scorer import Scorer, AggregateOutput
+from .scorer_registry import register_scorer
 
 DOMAIN_CAT2SUB_CAT = {
     "Art and Design": ["Art", "Art_Theory", "Design", "Music"],
@@ -410,6 +411,7 @@ def get_score(doc: Dataset, pred: str, random_choice: bool) -> int:
     return score
 
 
+@register_scorer("mmmu", "MMMU", "MMMUScorer")
 class MMMUScorer(Scorer):
     def score(self, refs: list[str], preds: list[str]) -> list[int]:
         docs = self.config.docs

src/eval_mm/metrics/rougel_scorer.py

@@ -6,6 +6,7 @@
 import emoji
 import unicodedata
 from .scorer import Scorer, AggregateOutput
+from .scorer_registry import register_scorer
 from concurrent.futures import ProcessPoolExecutor, Future
 
 
@@ -72,6 +73,7 @@ def rouge_ja(refs: list[str], preds: list[str]) -> dict:
     return {type: result[type].mid.fmeasure * 100 for type in rouge_types}
 
 
+@register_scorer("rougel", "RougeL", "RougeLScorer")
 class RougeLScorer(Scorer):
     @staticmethod
     def score(refs: list[str], preds: list[str]) -> list[float]: