Add new benchmark: Spanish bench (#2157)

* Add spanish_bench

* Add flores_es group

* Update _flores_common_yaml

* Delete lm_eval/tasks/spanish_bench/escola.yaml

* Delete escola from spanish_bench.yaml

* Delete escola from README.md

* pre-commit run --all-files

* Updated some task groupings and readme

---------

lm_eval/tasks/spanish_bench/flores_es/flores_pt-es.yaml

+# File generated by `create-yamls.py`
+include: _flores_common_yaml
+task: flores_pt-es
+doc_to_text: 'Portuguese sentence: {{sentence_por_Latn}}
+
+  Spanish sentence:'
+doc_to_target: '{{sentence_spa_Latn}}'

lm_eval/tasks/spanish_bench/mgsm_direct_es_spanish_bench.yaml

+include: ../mgsm/direct/mgsm_direct_es.yaml
+doc_to_target: '{{answer_number|string}}'
+doc_to_text: '{% if answer is not none %}{{question+"\nRespuesta: "}}{% else %}{{"Pregunta: "+question+"\nRespuesta: "}}{% endif %}'
+generation_kwargs:
+  until:
+    - "\n\n"
+    - "\n"
+
+task: mgsm_direct_es_spanish_bench

lm_eval/tasks/spanish_bench/mgsm_direct_es_v2.yaml

+include: ../mgsm/direct/mgsm_direct_es.yaml
+doc_to_target: '{{answer_number|string}}'
+doc_to_text: '{% if answer is not none %}{{question+"\nRespuesta: "}}{% else %}{{"Pregunta: "+question+"\nRespuesta: "}}{% endif %}'
+generation_kwargs:
+  until:
+    - "\n\n"
+    - "\n"
+
+task: mgsm_direct_es_v2

lm_eval/tasks/spanish_bench/paws_es.yaml

+group:
+    - pawsx
+task: paws_es
+dataset_path: paws-x
+dataset_name: es
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs_paraphrases
+doc_to_text: ''
+doc_to_target: label
+doc_to_choice: '{{[sentence1+", ¿verdad? No, "+sentence2, sentence1+", ¿verdad? Sí, "+sentence2]}}'
+target_delimiter: ''
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/spanish_bench/paws_es_spanish_bench.yaml

+task: paws_es_spanish_bench
+dataset_path: paws-x
+dataset_name: es
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+test_split: test
+process_docs: !function utils.process_docs_paraphrases
+doc_to_text: ''
+doc_to_target: label
+doc_to_choice: '{{[sentence1+", ¿verdad? No, "+sentence2, sentence1+", ¿verdad? Sí, "+sentence2]}}'
+target_delimiter: ''
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/spanish_bench/phrases_es/_phrases_es_common.yaml

+tag: phrases_es
+dataset_path: gplsi/ES-VA_translation_test
+output_type: generate_until
+training_split: null
+validation_split: null
+test_split: test
+fewshot_split: test
+num_fewshot: 5
+target_delimiter: ' '
+generation_kwargs:
+  until:
+    - "\n"
+metric_list:
+  - metric: bleu
+    aggregation: bleu
+    higher_is_better: true
+  - metric: ter
+    aggregation: ter
+    higher_is_better: false
+  - metric: chrf
+    aggregation: chrf
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/spanish_bench/phrases_es/phrases_es-va.yaml

+# File generated by `create-yamls.py`
+include: _phrases_es_common.yaml
+task: phrases_es-va
+doc_to_text: 'Oració en espanyol: {{es}}
+
+  Oració en valencià:'
+doc_to_target: '{{va}}'

lm_eval/tasks/spanish_bench/phrases_es/phrases_va-es.yaml

+# File generated by `create-yamls.py`
+include: _phrases_es_common.yaml
+task: phrases_va-es
+doc_to_text: 'Oració en valencià: {{va}}
+
+  Oració en espanyol:'
+doc_to_target: '{{es}}'

lm_eval/tasks/spanish_bench/spanish_bench.yaml

+group: spanish_bench
+task:
+  - belebele_spa_Latn
+  - wnli_es
+  - xnli_es_spanish_bench
+  - xstorycloze_es
+  - xquad_es
+  - xlsum_es
+  - paws_es_spanish_bench
+  - mgsm_direct_es_spanish_bench
+  - flores_es
+  - phrases_es
+metadata:
+  version: 1.0

lm_eval/tasks/spanish_bench/utils.py

+import re
+from itertools import product
+
+import evaluate
+import transformers.data.metrics.squad_metrics as squad_metrics
+
+from lm_eval.utils import general_detokenize
+
+
+def lowercase_first_letter(text):
+    return text[0].lower() + text[1:]
+
+
+def process_doc_nli(dataset):
+    def process_fn(doc):
+        # Detokenize(remove extra whitespaces)
+        doc["premise"] = general_detokenize(doc["premise"]).strip()
+        doc["hypothesis"] = general_detokenize(doc["hypothesis"]).strip()
+        # Remove last punctuation mark in the premise
+        doc["premise"] = (
+            doc["premise"][:-1]
+            if doc["premise"].endswith((".", ",", "!", "?"))
+            else doc["premise"]
+        )
+        # Lowercase the first letter in the hypothesis
+        doc["hypothesis"] = lowercase_first_letter(doc["hypothesis"])
+        # Ensure that the hypothesis ends with a dot
+        doc["hypothesis"] = (
+            (doc["hypothesis"] + ".")
+            if not doc["hypothesis"].endswith(".")
+            else doc["hypothesis"]
+        )
+        return doc
+
+    return dataset.map(process_fn)
+
+
+def process_results_qa(doc, results):
+    preds = results[0]
+    reference = doc["answers"]["text"][0]
+    # import code; code.interact(local=dict(globals(), **locals()))
+    f1_sum = squad_metrics.compute_f1(reference, preds)
+    exact_match = squad_metrics.compute_exact(reference, preds)
+    return {"f1": f1_sum, "exact_match": exact_match}
+
+
+def process_xlsum(dataset):
+    def _process_doc(doc):
+        # Remove double spaces
+        doc["text"] = re.sub(r" +", " ", doc["text"])
+        doc["summary"] = re.sub(r" +", " ", doc["summary"])
+        return doc
+
+    return dataset.map(_process_doc)
+
+
+def process_docs_paraphrases(dataset):
+    empty_docs = []
+
+    def _process_doc(doc):
+        if doc["sentence1"] not in [None, ""] and doc["sentence2"] not in [None, ""]:
+            doc["sentence1"] = general_detokenize(doc["sentence1"]).strip()
+            doc["sentence2"] = general_detokenize(doc["sentence2"]).strip()
+            # Remove final punctuation mark in the first sentence
+            if doc["sentence1"].endswith((".", ",", ";")):
+                doc["sentence1"] = doc["sentence1"][:-1]
+            # Start the second sentence in lowercase (to be used after "Yes, ...")
+            doc["sentence2"] = lowercase_first_letter(doc["sentence2"])
+            return doc
+        else:
+            empty_docs.append(doc)
+            return doc
+
+    if empty_docs != []:
+        len_empty_docs = len(empty_docs)
+        print(
+            f"Found {len_empty_docs} empty documents out of the {len(dataset)} total docs in the dataset: {empty_docs}"
+        )
+    return dataset.filter(
+        lambda doc: doc["sentence1"] not in [None, ""]
+        and doc["sentence2"] not in [None, ""]
+    ).map(_process_doc)
+
+
+def rouge1(items):
+    """
+    # passthrough for efficiency
+    """
+    return items
+
+
+def rouge1_agg(items):
+    """
+    Higher is better
+    """
+    refs = list(zip(*items))[0]
+    preds = list(zip(*items))[1]
+    rouge_scorer = evaluate.load("rouge")
+    # import code; code.interact(local=dict(globals(), **locals()))
+    return rouge_scorer.compute(predictions=preds, references=refs)["rouge1"]

lm_eval/tasks/spanish_bench/wnli_es.yaml

+task: wnli_es
+dataset_path: PlanTL-GOB-ES/wnli-es
+dataset_name: null
+output_type: multiple_choice
+training_split: train
+validation_split: validation
+test_split: null
+doc_to_text: "{{sentence1}}\nPregunta: {{sentence2}} ¿Verdadero o Falso?\nRespuesta:"
+doc_to_target: label
+doc_to_choice: ["Falso", "Verdadero"]
+metric_list:
+  - metric: acc
+metadata:
+  version: 1.0

lm_eval/tasks/spanish_bench/xlsum_es.yaml

+task: xlsum_es
+dataset_path: csebuetnlp/xlsum
+dataset_name: spanish
+doc_to_text: 'Texto: {{text}}
+
+  Resumen:'
+doc_to_target: '{{summary}}'
+output_type: generate_until
+test_split: test
+training_split: train
+validation_split: validation
+fewshot_split: train
+process_docs: !function utils.process_xlsum
+metric_list:
+  - metric: bleu
+    aggregation: bleu
+    higher_is_better: true
+  - metric: !function utils.rouge1
+    aggregation: !function utils.rouge1_agg
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/spanish_bench/xnli_es.yaml

+# Task configuration derived from Eleuther AI's implementation as of March 22, 2024, supplemented with an additional preprocessing function
+task: xnli_es
+dataset_path: xnli
+dataset_name: es
+output_type: multiple_choice
+doc_to_choice: '{{[premise+", ¿correcto? Sí, "+hypothesis,premise+", ¿correcto? Así
+  que, "+hypothesis,premise+", ¿correcto? No, "+hypothesis]}}'
+doc_to_text: ''
+target_delimiter: ''
+process_docs: !function utils.process_doc_nli
+training_split: null
+validation_split: validation
+doc_to_target: label
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/spanish_bench/xnli_es_spanish_bench.yaml

+# Task configuration derived from Eleuther AI's implementation as of March 22, 2024, supplemented with an additional preprocessing function
+task: xnli_es_spanish_bench
+dataset_path: xnli
+dataset_name: es
+output_type: multiple_choice
+doc_to_choice: '{{[premise+", ¿correcto? Sí, "+hypothesis,premise+", ¿correcto? Así
+  que, "+hypothesis,premise+", ¿correcto? No, "+hypothesis]}}'
+doc_to_text: ''
+target_delimiter: ''
+process_docs: !function utils.process_doc_nli
+training_split: null
+validation_split: validation
+doc_to_target: label
+metric_list:
+  - metric: acc
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0

lm_eval/tasks/spanish_bench/xquad_es.yaml

+task: xquad_es
+dataset_path: xquad
+dataset_name: xquad.es
+output_type: generate_until
+doc_to_text: "Contexto: {{context}}\n\nPregunta: {{question}}\n\nRespuesta:"
+doc_to_target: '{{answers["text"][0]}}'
+validation_split: validation
+target_delimiter: ' '
+process_results: !function utils.process_results_qa
+generation_kwargs:
+  until:
+    - "\n"
+  do_sample: false
+  temperature: 0.0
+metric_list:
+  - metric: exact_match
+    aggregation: mean
+    higher_is_better: true
+  - metric: f1
+    aggregation: mean
+    higher_is_better: true
+metadata:
+  version: 1.0