Xp3zhmt

promptsource/machine_translate.py

@@ -3,96 +3,47 @@
 
 from promptsource.templates import Template, TemplateCollection
 
-
-### XNLI
-
-PROMPTS = [
-    "GPT-3 style",
-    "can we infer",
-    "justified in saying",
-    "guaranteed/possible/impossible",
-    "MNLI crowdsource",
-]
-
-LANGS = [
-    "ar",
-    "es",
-    "fr",
-    "hi",
-    "sw",
-    "ur",
-    "vi",
-    "zh",
-]
-
-SOURCE_DATASET = TARGET_DATASET = "xnli"
-SOURCE_LANG = "en"
-
-
-### XCOPA
-
-PROMPTS = [
-    "best_option",
-    'C1 or C2? premise, so/because…',
-    "i_am_hesitating",
-    "cause_effect",
-    "plausible_alternatives",
-]
-
-LANGS = [
-    "id",
-    "sw",
-    "ta",
-    "vi",
-    "zh",
-]
-
-SOURCE_DATASET = "super_glue/copa"
-SOURCE_LANG = None
-TARGET_DATASET = "xcopa"
-
-### XSTORY_CLOZE
-
-PROMPTS = [
-    "Answer Given options",
-    'Choose Story Ending',
-    "Story Continuation and Options",
-    "Generate Ending",
-    "Novel Correct Ending",
-]
-
-LANGS = [
-    "ar",
-    "es",
-    "eu",
-    "hi",
-    "id",
-    "zh",
+DS_TO_ENG_PROMPT = {
+    "xcopa": "en",
+    "Muennighoff/xstory_cloze": "en",
+    "Muennighoff/xwinograd": "en",
+    'GEM/wiki_lingua': 'en_en', # Contains correct language names
+    'xnli': 'en',
+    "paws-x": "en",
+    "mlqa": "mlqa.en.en",
+    "xquad": "xquad.en",
+    "khalidalt/tydiqa-primary": "english",
+    "khalidalt/tydiqa-goldp": "english",
+    "pasinit/xlwic": "en",
+    "GEM/xlsum": "english",
+    "GEM/BiSECT": "en",
+}
+
+### ZH Datasets
+
+DATASETS = [
+    ('xquad', 'xquad.zh'),
+    # Context & Answer is in ZH
+    ('mlqa', 'mlqa.zh.ar'),
+    ('mlqa', 'mlqa.zh.vi'),
+    ('mlqa', 'mlqa.zh.es'),
+    ('mlqa', 'mlqa.zh.en'),
+    ('mlqa', 'mlqa.zh.hi'),
+    ('paws-x', 'zh'),
+    ('clue', 'c3'),
+    ('clue', 'cmrc2018'),
+    ('clue', 'csl'),
+    ('clue', 'drcd'),
+    ('clue', 'tnews'),
+    ('pasinit/xlwic', "xlwic_en_zh"),
+    ('GEM/xlsum', "chinese_simplified"),
+    # ('GEM/xlsum', "chinese_traditional"),
+    # For WikiLingua there are already ZH prompts (except for xp3long prompts)
+    ("xquad", )
 ]
 
-SOURCE_DATASET = TARGET_DATASET = "Muennighoff/xstory_cloze"
-SOURCE_LANG = "en"
-
-### XWINOGRAD
-
-PROMPTS = [
-    "Replace",
-    "stand for",
-    "True or False",
-    "does underscore refer to",
-    "underscore refer to",
-]
-
-LANGS = [
-    "fr",
-    "pt",
-    "zh",
-]
-
-SOURCE_DATASET = TARGET_DATASET = "Muennighoff/xwinograd"
-SOURCE_LANG = "en"
-
 
+LANG = "zh"
 
 # Path to key
 os.environ["GOOGLE_APPLICATION_CREDENTIALS"] = "/Users/niklasmuennighoff/Desktop/gcp_translate_key.json"
@@ -147,22 +98,28 @@ def normalize_string(zh_string, en_string):
 
 
 template_collection = TemplateCollection()
-source_templates = template_collection.get_dataset(SOURCE_DATASET, SOURCE_LANG)
 
-for lang in LANGS:
-    target_templates = template_collection.get_dataset(TARGET_DATASET, lang)
+for (ds_name, subset_name) in DATASETS:
+
+    subset_name_eng = subset_name
+    if ds_name in DS_TO_ENG_PROMPT:
+        subset_name_eng = DS_TO_ENG_PROMPT[ds_name]
+
+    source_templates = template_collection.get_dataset(ds_name, subset_name_eng)
+    #for lang in LANGS:
+    target_templates = template_collection.get_dataset(ds_name, subset_name)
     for uid, template in source_templates.templates.items():
-        if template.name.strip() not in PROMPTS:
+        if not("xp3long" in template.name.strip()):# not in PROMPTS:
             continue
-        print(f"Translating {template.name.strip()} to {lang}")
+        print(f"Translating {template.name.strip()} to {LANG}")
         answer_choices = []
         if template.answer_choices is not None:
             choices = template.answer_choices.split("|||")
             for c in choices:
-                answer_choices.append(normalize_string(translate(lang, c.strip()), c.strip()))
+                answer_choices.append(normalize_string(translate(LANG, c.strip()), c.strip()))
         or_jinja = template.jinja.strip()
-        jinja = normalize_string(translate(lang, or_jinja), or_jinja)
-        template_name = template.name.strip() + f"_{lang}mt"
+        jinja = normalize_string(translate(LANG, or_jinja), or_jinja)
+        template_name = template.name.strip() + f"_{LANG}mt"
         target_template = Template(
             template_name, jinja=jinja, reference="", answer_choices=" ||| ".join(answer_choices)
         )

promptsource/templates.py

@@ -841,6 +841,7 @@ def read_from_file(self) -> Dict:
                 "Please ignore this warning if you are creating new prompts for this dataset."
             )
             return {}
+        print(self.dataset_name, self.yaml_path)
         yaml_dict = yaml.load(open(self.yaml_path, "r"), Loader=yaml.FullLoader)
         return yaml_dict[self.TEMPLATES_KEY]
 

promptsource/templates/GEM/wiki_lingua/zh/templates.yaml

@@ -71,3 +71,30 @@ templates:
       original_task: true
     name: summarize_above_zh
     reference: xsum templates
+  dfa7b514-7385-4855-bb90-253073a34fde: !Template
+    answer_choices: null
+    id: dfa7b514-7385-4855-bb90-253073a34fde
+    jinja: "{{target}}\n\n鉴于上述总结，为它写一个详细的文本。||| {{source}}"
+    metadata: !TemplateMetadata
+      choices_in_prompt: false
+      languages: []
+      metrics:
+      - ROUGE
+      - BLEU
+      original_task: true
+    name: xp3longwritearticle_zhmt
+    reference: ''
+  dff8b414-7485-4855-bb90-253073a34fde: !Template
+    answer_choices: null
+    id: dff8b414-7485-4855-bb90-253073a34fde
+    jinja: "{{target}}\n\n我对此很感兴趣，但我只有几分钟的时间。
+    你能不能给我最多前500个字符的详细解释关于那个？ ||| {{source[:500]}}"
+    metadata: !TemplateMetadata
+      choices_in_prompt: false
+      languages: []
+      metrics:
+      - ROUGE
+      - BLEU
+      original_task: true
+    name: xp3longchars_zhmt
+    reference: ''

promptsource/templates/GEM/xlsum/chinese_simplified/templates.yaml

@@ -0,0 +1,115 @@
+dataset: GEM/xlsum
+subset: chinese_simplified
+templates:
+  0bf9d87c-c6ac-4ff3-a956-e57ec434c4df: !Template
+    answer_choices: ''
+    id: 0bf9d87c-c6ac-4ff3-a956-e57ec434c4df
+    jinja: 标题：{{title}}\n给定上面虚构文章的标题，想象这篇文章。\n ||| {{text[:7000]}}
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: imaginearticle_zhmt
+    reference: ''
+  70151209-fdd4-4b78-988d-629ff5da1239: !Template
+    answer_choices: ''
+    id: 70151209-fdd4-4b78-988d-629ff5da1239
+    jinja: '{{text[:1000]}}... 继续这篇文章，最多 4000 个字符：||| {{text[1000:5000]}}'
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: xp3longcontinue_zhmt
+    reference: ''
+  900008a7-0cde-4f0d-8fd3-2ff9cb5b36f1: !Template
+    answer_choices: ''
+    id: 900008a7-0cde-4f0d-8fd3-2ff9cb5b36f1
+    jinja: 给定以下文章的标题和摘要，生成一篇短篇文章或一篇长篇文章的开头以配合它们。标题：{{title}}\n摘要：{{target}}\n文章（最多
+      500 个字符）：||| {{text[:500]}}
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: xp3longgenarticle_zhmt
+    reference: ''
+  9387daa0-a014-4df0-9d21-ff086dc5df4a: !Template
+    answer_choices: ''
+    id: 9387daa0-a014-4df0-9d21-ff086dc5df4a
+    jinja: 要总结的文档：{{text[:8500]}}\n以与文档相同的语言进行总结：||| {{target}}
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: docsummary_zhmt
+    reference: ''
+  97e1cdb0-1c5d-4421-a15b-faf02cd812a8: !Template
+    answer_choices: ''
+    id: 97e1cdb0-1c5d-4421-a15b-faf02cd812a8
+    jinja: '{{text}} \n\n给我上面文章的好标题。 ||| {{title}}'
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: goodtitle_zhmt
+    reference: ''
+  9d46d594-7463-4f74-bfd9-b5cbf804e94b: !Template
+    answer_choices: ''
+    id: 9d46d594-7463-4f74-bfd9-b5cbf804e94b
+    jinja: 标题：{{title}}\n给定上面虚构文章的标题，想象这篇文章。\n ||| {{text[:7000]}}
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: xp3longimaginearticle_zhmt
+    reference: ''
+  bfd429f5-d75d-46a2-8224-874d1109c310: !Template
+    answer_choices: ''
+    id: bfd429f5-d75d-46a2-8224-874d1109c310
+    jinja: '{{title}}\n{{text[:5000]}}\n\ntl;博士：||| {{target}}'
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: tldr_zhmt
+    reference: ''
+  d4f2980a-cb44-4e47-b6c3-3df858c82126: !Template
+    answer_choices: ''
+    id: d4f2980a-cb44-4e47-b6c3-3df858c82126
+    jinja: 给定以下文章的标题和摘要，生成一篇短篇文章或一篇长篇文章的开头以配合它们。标题：{{title}}\n摘要：{{target}}\n文章（最多
+      500 个字符）：||| {{text[:500]}}
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: genarticle_zhmt
+    reference: ''
+  dca03c69-940f-46b0-acd8-16bbcbd6bec9: !Template
+    answer_choices: ''
+    id: dca03c69-940f-46b0-acd8-16bbcbd6bec9
+    jinja: '...{{text[3000:3500]}}... 写下文章的其余部分：||| {{text[5000:]}}'
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: xp3longrest_zhmt
+    reference: ''
+  ff7e01a6-9c23-4228-85c8-b84432deeb50: !Template
+    answer_choices: ''
+    id: ff7e01a6-9c23-4228-85c8-b84432deeb50
+    jinja: 内容：{{text[:7000]}}\n前面的内容可以总结如下：||| {{target}}
+    metadata: !TemplateMetadata
+      choices_in_prompt: null
+      languages: null
+      metrics: null
+      original_task: null
+    name: prevcontent_zhmt
+    reference: ''