Add MultiFin high-level v0 datasets, requests, and manifests

datasets/multifin/highlevel/v0/clean_meta.json

@@ -0,0 +1,63 @@
+{
+  "dataset_id": "multifin_highlevel_v0",
+  "config": "all_languages_highlevel",
+  "source_dataset": "awinml/MultiFin",
+  "splits": {
+    "train": {
+      "n_in": 6430,
+      "n_out": 6430,
+      "label_counts": {
+        "Tax & Accounting": 2186,
+        "Government & Controls": 597,
+        "Business & Management": 1249,
+        "Industry": 770,
+        "Finance": 941,
+        "Technology": 687
+      },
+      "out_file": "data/multifin/processed/multifin_highlevel_train_clean.jsonl"
+    },
+    "validation": {
+      "n_in": 1608,
+      "n_out": 1608,
+      "label_counts": {
+        "Tax & Accounting": 525,
+        "Technology": 180,
+        "Business & Management": 340,
+        "Government & Controls": 149,
+        "Finance": 214,
+        "Industry": 200
+      },
+      "out_file": "data/multifin/processed/multifin_highlevel_validation_clean.jsonl"
+    },
+    "test": {
+      "n_in": 2010,
+      "n_out": 2010,
+      "label_counts": {
+        "Finance": 292,
+        "Technology": 221,
+        "Tax & Accounting": 660,
+        "Government & Controls": 166,
+        "Business & Management": 402,
+        "Industry": 269
+      },
+      "out_file": "data/multifin/processed/multifin_highlevel_test_clean.jsonl"
+    }
+  },
+  "overall_label_counts": {
+    "Tax & Accounting": 3371,
+    "Government & Controls": 912,
+    "Business & Management": 1991,
+    "Industry": 1239,
+    "Finance": 1447,
+    "Technology": 1088
+  },
+  "allowed_labels": [
+    "Business & Management",
+    "Finance",
+    "Government & Controls",
+    "Industry",
+    "Tax & Accounting",
+    "Technology"
+  ],
+  "split_policy": "preserve_upstream_train_validation_test"
+}

manifests/datasets/multifin_highlevel_v0/checksums.sha256

@@ -0,0 +1,4 @@
+a0517cf8461b517ab5d84adf4f62988f0d8aad75fd92a918e61b18f6f66d41bd  data/multifin/processed/multifin_highlevel_train_clean.jsonl
+fde64013f698f4067d9a2a97ebc631c23f0571494080e9cead3aac7e4bed0bb0  data/multifin/processed/multifin_highlevel_validation_clean.jsonl
+f5f515a3ea7ea7b1035fc48f7871931df34ce768e754eb9ce97630dbbdf12fc4  data/multifin/processed/multifin_highlevel_test_clean.jsonl
+d85fbb55759f4411e5e6ec87f5335a57ce0a513bf50b4843ace0a9c8e66e658e  data/multifin/processed/multifin_highlevel_clean_meta.json

manifests/datasets/multifin_highlevel_v0/dataset_spec.json

@@ -0,0 +1,60 @@
+{
+  "dataset_id": "multifin_highlevel_v0",
+  "dataset_name": "MultiFin High-Level Topic Classification",
+  "version": "0.1.0",
+  "dataset_folder": "datasets/multifin_highlevel_v0",
+  "source": {
+    "type": "huggingface",
+    "hf_dataset_id": "awinml/MultiFin",
+    "hf_config": "all_languages_highlevel"
+  },
+  "files": {
+    "raw_files": {
+      "train": "data/multifin/raw/multifin_highlevel_train_raw.jsonl",
+      "validation": "data/multifin/raw/multifin_highlevel_validation_raw.jsonl",
+      "test": "data/multifin/raw/multifin_highlevel_test_raw.jsonl"
+    },
+    "raw_meta_file": "data/multifin/raw/multifin_highlevel_raw_meta.json",
+    "processed_files": {
+      "train": "data/multifin/processed/multifin_highlevel_train_clean.jsonl",
+      "validation": "data/multifin/processed/multifin_highlevel_validation_clean.jsonl",
+      "test": "data/multifin/processed/multifin_highlevel_test_clean.jsonl"
+    },
+    "processed_meta_file": "data/multifin/processed/multifin_highlevel_clean_meta.json"
+  },
+  "label_scheme": {
+    "type": "single_label_multiclass",
+    "task": "topic_classification",
+    "classes": [
+      "Technology",
+      "Industry",
+      "Tax & Accounting",
+      "Finance",
+      "Government & Controls",
+      "Business & Management"
+    ]
+  },
+  "split_policy": {
+    "type": "preserve_upstream",
+    "splits": [
+      "train",
+      "validation",
+      "test"
+    ]
+  },
+  "hf_publish": {
+    "repo_id": "mmmikolajczak/st312-data",
+    "repo_type": "dataset",
+    "dataset_root": "datasets/multifin/highlevel/v0",
+    "upstream_license": "cc-by-sa-4.0",
+    "upstream_license_name": "Creative Commons Attribution Share Alike 4.0 International",
+    "upstream_license_url": "https://creativecommons.org/licenses/by-sa/4.0/",
+    "upstream_license_note": "License listed on the Hugging Face dataset card for awinml/MultiFin."
+  },
+  "status": {
+    "raw_ingested": true,
+    "cleaned": true,
+    "requests_built": true,
+    "published_to_hf": false
+  }
+}

manifests/tasks/multifin_highlevel_topic_v0/task_spec.json

@@ -0,0 +1,36 @@
+{
+  "task_id": "TA_TOPIC_MULTIFIN_HIGHLEVEL_v0",
+  "task_name": "MultiFin High-Level Topic Classification",
+  "version": "0.1.0",
+  "task_type": "single_label_multiclass",
+  "dataset": {
+    "dataset_id": "multifin_highlevel_v0",
+    "train_file": "data/multifin/processed/multifin_highlevel_train_clean.jsonl",
+    "validation_file": "data/multifin/processed/multifin_highlevel_validation_clean.jsonl",
+    "test_file": "data/multifin/processed/multifin_highlevel_test_clean.jsonl"
+  },
+  "input_schema": {
+    "required_fields": ["data.text"],
+    "optional_fields": ["data.language"]
+  },
+  "output_schema": {
+    "type": "json_object",
+    "strict_single_key": "topic",
+    "allowed_values": [
+      "Technology",
+      "Industry",
+      "Tax & Accounting",
+      "Finance",
+      "Government & Controls",
+      "Business & Management"
+    ]
+  },
+  "prompt_template": {
+    "system": "You are a financial topic classifier. Return ONLY a JSON object with exactly one key: \"topic\". The value must be exactly one of the following labels: \"Technology\", \"Industry\", \"Tax & Accounting\", \"Finance\", \"Government & Controls\", \"Business & Management\". Do not include any extra text.",
+    "user": "Classify the following financial text into exactly one high-level topic category.\n\nText: {{data.text}}"
+  },
+  "metrics": [
+    "accuracy",
+    "macro_f1"
+  ]
+}

tasks/multifin_highlevel_topic_v0/README.md

@@ -0,0 +1,35 @@
+# MultiFin High-Level Topic Task (v0)
+
+Task module for 6-way topic classification on `multifin_highlevel_v0`.
+
+## Task ID
+`TA_TOPIC_MULTIFIN_HIGHLEVEL_v0`
+
+## Dataset module
+- Dataset ID: `multifin_highlevel_v0`
+- Train: `data/multifin/processed/multifin_highlevel_train_clean.jsonl`
+- Validation: `data/multifin/processed/multifin_highlevel_validation_clean.jsonl`
+- Test: `data/multifin/processed/multifin_highlevel_test_clean.jsonl`
+
+## Labels
+- `Technology`
+- `Industry`
+- `Tax & Accounting`
+- `Finance`
+- `Government & Controls`
+- `Business & Management`
+
+## Output schema (strict)
+Return ONLY:
+```json
+{"topic": "Finance"}
+```
+
+Rules:
+- exactly one key: `topic`
+- value must be one of the six allowed labels
+- no extra keys
+- no extra text
+
+## Scripts
+- Reward parser: `scripts/tasks/multifin_highlevel_topic_v0/reward_multifin_highlevel.py`