feat: chat agent schema

seacrowd/sea_datasets/onto4all/onto4all.py

@@ -52,7 +52,7 @@
 
 _URLS = "https://huggingface.co/datasets/ontocord/onto4all/resolve/main/data/train-00000-of-00001.parquet?download=true"
 
-_SUPPORTED_TASKS = [Tasks.QUESTION_ANSWERING]
+_SUPPORTED_TASKS = [Tasks.MULTI_TURN_CONVERSATION]
 
 _SOURCE_VERSION = "1.0.0"
 
@@ -73,10 +73,10 @@ class Onto4AllDataset(datasets.GeneratorBasedBuilder):
             subset_id=f"{_DATASETNAME}",
         ),
         SEACrowdConfig(
-            name=f"{_DATASETNAME}_seacrowd_qa",
+            name=f"{_DATASETNAME}_seacrowd_chat",
             version=SEACROWD_VERSION,
             description=f"{_DATASETNAME} SEACrowd schema",
-            schema="seacrowd_qa",
+            schema="seacrowd_chat",
             subset_id=f"{_DATASETNAME}",
         ),
     ]
@@ -98,8 +98,11 @@ def _info(self) -> datasets.DatasetInfo:
                 }
             )
 
-        elif self.config.schema == "seacrowd_qa":
-            features = schemas.qa_features
+        elif self.config.schema == "seacrowd_chat":
+            features = schemas.chat_features
+            features["meta"] = {
+                "type": datasets.Value("string")
+            }
 
         return datasets.DatasetInfo(
             description=_DESCRIPTION,
@@ -140,9 +143,8 @@ def _generate_examples(self, filepath: Path) -> Tuple[int, Dict]:
                     "type": row["type"],
                     "conversation": conversation,
                 }
-                break
 
-        elif self.config.schema == "seacrowd_qa":
+        elif self.config.schema == "seacrowd_chat":
             for i, row in df.iterrows():
                 context = ""
                 question = ""
@@ -158,12 +160,18 @@ def _generate_examples(self, filepath: Path) -> Tuple[int, Dict]:
 
                 yield i, {
                     "id": row["id"],
-                    "question_id": row["id"],
-                    "document_id": "",
-                    "question": question,
-                    "type": row["type"],
-                    "choices": [],
-                    "context": context,
-                    "answer": [answer],
-                    "meta": {},
+                    "input": [
+                        {
+                            "role": "system",
+                            "content": context,
+                        },
+                        {
+                            "role": "user",
+                            "content": question,
+                        },
+                    ],
+                    "output": answer,
+                    "meta": {
+                        "type": row["type"],
+                    },
                 }

seacrowd/utils/constants.py

@@ -10,6 +10,7 @@
     pairs_features_score,
     pairs_multi_features,
     qa_features,
+    chat_features,
     image_features,
     image_multi_features,
     imqa_features,
@@ -105,6 +106,7 @@ class Tasks(Enum):
     # Multi Text Generation
     DIALOGUE_SYSTEM = "DS"
     E2E_TASK_ORIENTED_DIALOGUE = "TOD"
+    MULTI_TURN_CONVERSATION = "MTC"
 
     # Self Supervised & Unsupervised Text
     PROMPTING = "PRT"
@@ -246,6 +248,7 @@ class Licenses(Enum):
     Tasks.TOKEN_LEVEL_LANGUAGE_IDENTIFICATION: "SEQ_LABEL",
     Tasks.COMMONSENSE_REASONING: "QA",
     Tasks.QUESTION_ANSWERING: "QA",
+    Tasks.MULTI_TURN_CONVERSATION: "CHAT",
     Tasks.CONCEPT_ALIGNMENT_CLASSIFICATION: "PAIRS",
     Tasks.NEXT_SENTENCE_PREDICTION: "PAIRS",
     Tasks.TEXT_RETRIEVAL: "PAIRS",
@@ -313,6 +316,7 @@ class Licenses(Enum):
     "KB": kb_features,
     "TREE": tree_features,
     "QA": qa_features,
+    "CHAT": chat_features,
     "T2T": text2text_features,
     "TEXT": text_features(),
     "TEXT_MULTI": text_multi_features(),

seacrowd/utils/schemas/__init__.py

@@ -5,6 +5,7 @@
 from .pairs import features_with_continuous_label as pairs_features_score
 from .pairs_multilabel import features as pairs_multi_features
 from .qa import features as qa_features
+from .chat import features as chat_features
 from .image import features as image_features
 from .image import multi_features as image_multi_features
 from .imqa import features as imqa_features
@@ -28,6 +29,7 @@
     "pairs_features_score",
     "pairs_multi_features",
     "qa_features",
+    "chat_features",
     "image_features",
     "image_multi_features",
     "imqa_features",

seacrowd/utils/schemas/chat.py

@@ -0,0 +1,24 @@
+"""
+Conversational Chat Schema
+"""
+import datasets
+
+features = datasets.Features(
+    {
+        "id": datasets.Value("string"),
+        "input": datasets.Sequence({
+            "role": datasets.ClassLabel(names=["system", "user", "assistant"]),
+            "content": datasets.Value("string"),
+        }),
+        "output": datasets.Value("string"),
+
+        # the schema of 'meta' aren't specified either to allow some flexibility
+        "meta": {}
+
+        # notes on how to use this field of 'meta'
+        # you can choose two of options:
+        # 1. defining as empty dict if you don't think it's usable in `_generate_examples`, or
+        # 2. defining meta as dict of key with intended colname meta and its val with dataset.Features class
+        #    in `_info` Dataloader method then populate it with the values in `_general_examples` Dataloader method
+    }
+)