Add FinBen FINER-ORD v0 datasets, requests, and manifests

datasets/finer_ord/finben/v0/clean_meta.json

@@ -0,0 +1,35 @@
+{
+  "dataset_id": "finben_finer_ord_v0",
+  "config": "token_labels_v0",
+  "source_dataset": "TheFinAI/finben-finer-ord",
+  "n_in": 1075,
+  "n_out": 1075,
+  "allowed_labels": [
+    "B-LOC",
+    "B-ORG",
+    "B-PER",
+    "I-LOC",
+    "I-ORG",
+    "I-PER",
+    "O"
+  ],
+  "token_label_counts": {
+    "O": 24130,
+    "B-PER": 284,
+    "I-PER": 182,
+    "B-LOC": 300,
+    "I-LOC": 128,
+    "B-ORG": 544,
+    "I-ORG": 389
+  },
+  "sentence_signature_counts": {
+    "NONE": 524,
+    "LOC+PER": 35,
+    "LOC": 77,
+    "PER": 107,
+    "ORG": 200,
+    "LOC+ORG": 59,
+    "ORG+PER": 50,
+    "LOC+ORG+PER": 23
+  }
+}

datasets/finer_ord/finben/v0/split_meta.json

@@ -0,0 +1,95 @@
+{
+  "dataset_id": "finben_finer_ord_v0",
+  "config": "stratified_80_20_v0",
+  "source_clean_file": "data/finer_ord/processed/finben_finer_ord_all_clean.jsonl",
+  "seed": 42,
+  "test_fraction": 0.2,
+  "n_total": 1075,
+  "n_train": 860,
+  "n_test": 215,
+  "stratify_on": "meta.label_signature",
+  "per_signature_split": {
+    "LOC": {
+      "total": 77,
+      "train": 62,
+      "test": 15
+    },
+    "LOC+ORG": {
+      "total": 59,
+      "train": 47,
+      "test": 12
+    },
+    "LOC+ORG+PER": {
+      "total": 23,
+      "train": 18,
+      "test": 5
+    },
+    "LOC+PER": {
+      "total": 35,
+      "train": 28,
+      "test": 7
+    },
+    "NONE": {
+      "total": 524,
+      "train": 419,
+      "test": 105
+    },
+    "ORG": {
+      "total": 200,
+      "train": 160,
+      "test": 40
+    },
+    "ORG+PER": {
+      "total": 50,
+      "train": 40,
+      "test": 10
+    },
+    "PER": {
+      "total": 107,
+      "train": 86,
+      "test": 21
+    }
+  },
+  "sentence_signature_counts_train": {
+    "NONE": 419,
+    "LOC+ORG": 47,
+    "ORG": 160,
+    "LOC": 62,
+    "LOC+PER": 28,
+    "PER": 86,
+    "ORG+PER": 40,
+    "LOC+ORG+PER": 18
+  },
+  "sentence_signature_counts_test": {
+    "NONE": 105,
+    "ORG": 40,
+    "PER": 21,
+    "ORG+PER": 10,
+    "LOC+ORG": 12,
+    "LOC+PER": 7,
+    "LOC": 15,
+    "LOC+ORG+PER": 5
+  },
+  "token_label_counts_train": {
+    "O": 19339,
+    "B-ORG": 436,
+    "B-LOC": 233,
+    "I-ORG": 320,
+    "B-PER": 226,
+    "I-PER": 144,
+    "I-LOC": 109
+  },
+  "token_label_counts_test": {
+    "O": 4791,
+    "B-ORG": 108,
+    "B-PER": 58,
+    "I-PER": 38,
+    "I-ORG": 69,
+    "B-LOC": 67,
+    "I-LOC": 19
+  },
+  "files": {
+    "train": "data/finer_ord/processed/finben_finer_ord_train.jsonl",
+    "test": "data/finer_ord/processed/finben_finer_ord_test.jsonl"
+  }
+}

manifests/datasets/finben_finer_ord_v0/checksums.sha256

@@ -0,0 +1,5 @@
+60029d0a67c05df8c2c5692a06e2cf3993583a48132b6fd2ef70aafc83836cd6  data/finer_ord/processed/finben_finer_ord_all_clean.jsonl
+c81275976958c7b1c9518e8e39cd98261379a92ac8e7912b2bfd7e3e0cd383c5  data/finer_ord/processed/finben_finer_ord_train.jsonl
+7a8e34557a9ef4d84a35f38b06f22b6f569804f26dfa6cf732f39a073ba857b2  data/finer_ord/processed/finben_finer_ord_test.jsonl
+6ec852a13d2e2b0cb93b9ffd86c62c38c4984f15779fcb116919fad8803c0897  data/finer_ord/processed/finben_finer_ord_clean_meta.json
+6a2ce8a0a5ebc559e183cc13d2b9cd4f230f6b0ab1341ec5fafeb4e2041f210d  data/finer_ord/processed/finben_finer_ord_split_meta.json

manifests/datasets/finben_finer_ord_v0/dataset_spec.json

@@ -0,0 +1,64 @@
+{
+  "dataset_id": "finben_finer_ord_v0",
+  "dataset_name": "FinBen FINER-ORD",
+  "version": "0.1.0",
+  "dataset_folder": "datasets/finben_finer_ord_v0",
+  "provenance": {
+    "upstream_source": "TheFinAI/finben-finer-ord",
+    "source_kind": "finben_mirror",
+    "finben_reference": "TheFinAI/finben-finer-ord"
+  },
+  "processed": {
+    "config": "stratified_80_20_v0",
+    "files": {
+      "all_clean": "data/finer_ord/processed/finben_finer_ord_all_clean.jsonl",
+      "train": "data/finer_ord/processed/finben_finer_ord_train.jsonl",
+      "test": "data/finer_ord/processed/finben_finer_ord_test.jsonl"
+    },
+    "meta_files": {
+      "clean_meta": "data/finer_ord/processed/finben_finer_ord_clean_meta.json",
+      "split_meta": "data/finer_ord/processed/finben_finer_ord_split_meta.json"
+    }
+  },
+  "label_scheme": {
+    "type": "token_classification_bio",
+    "task": "named_entity_recognition",
+    "entities": [
+      "PER",
+      "LOC",
+      "ORG"
+    ],
+    "classes": [
+      "O",
+      "B-PER",
+      "I-PER",
+      "B-LOC",
+      "I-LOC",
+      "B-ORG",
+      "I-ORG"
+    ]
+  },
+  "split_policy": {
+    "type": "local_stratified_train_test",
+    "train_fraction": 0.8,
+    "test_fraction": 0.2,
+    "stratify_on": "meta.label_signature",
+    "seed": 42
+  },
+  "hf_publish": {
+    "repo_id": "mmmikolajczak/st312-data",
+    "repo_type": "dataset",
+    "dataset_root": "datasets/finer_ord/finben/v0",
+    "upstream_license": "cc-by-nc-4.0",
+    "upstream_license_name": "Creative Commons Attribution-NonCommercial 4.0 International",
+    "upstream_license_url": "https://creativecommons.org/licenses/by-nc/4.0/",
+    "upstream_license_note": "License listed on the Hugging Face dataset card for TheFinAI/finben-finer-ord (verify when comparing with original source)."
+  },
+  "status": {
+    "raw_ingested": true,
+    "cleaned": true,
+    "split": true,
+    "requests_built": true,
+    "published_to_hf": false
+  }
+}

manifests/tasks/finben_finer_ord_ner_v0/task_spec.json

@@ -0,0 +1,64 @@
+{
+  "task_id": "TA_NER_FINER_ORD_FINBEN_v0",
+  "task_name": "FinBen FINER-ORD NER (BIO token tagging)",
+  "version": "0.1.0",
+  "task_type": "token_classification",
+  "dataset": {
+    "dataset_id": "finben_finer_ord_v0",
+    "train_file": "data/finer_ord/processed/finben_finer_ord_train.jsonl",
+    "test_file": "data/finer_ord/processed/finben_finer_ord_test.jsonl"
+  },
+  "input_schema": {
+    "required_fields": [
+      "data.tokens"
+    ]
+  },
+  "output_schema": {
+    "type": "json_object",
+    "strict_single_key": "tags",
+    "allowed_tag_values": [
+      "O",
+      "B-PER",
+      "I-PER",
+      "B-LOC",
+      "I-LOC",
+      "B-ORG",
+      "I-ORG"
+    ]
+  },
+  "prompt_template": {
+    "system": "You are a financial NER tagger. Return ONLY a JSON object with exactly one key: \"tags\". The value must be a JSON array of BIO tags (same length as input tokens). Allowed tags: O, B-PER, I-PER, B-LOC, I-LOC, B-ORG, I-ORG. Do not include any extra text.",
+    "user": "Assign a BIO NER tag to each token in order. Output JSON only.\n\nTokens (one per line):\n{{data.tokens_lines}}"
+  },
+  "metrics": [
+    "exact_sequence_accuracy",
+    "token_accuracy",
+    "entity_span_precision",
+    "entity_span_recall",
+    "entity_span_f1",
+    "micro_f1_non_o"
+  ],
+  "reward": {
+    "script": "scripts/tasks/finben_finer_ord_ner_v0/reward_finer_ord_ner.py",
+    "format_reward": {
+      "type": "strict_json_schema",
+      "weight": 1.0
+    },
+    "correctness_reward": {
+      "name": "shaped_ner_reward_v1",
+      "range": [
+        0.0,
+        1.0
+      ],
+      "components": {
+        "entity_span_f1": 0.5,
+        "token_accuracy": 0.4,
+        "exact_sequence_match": 0.1
+      }
+    },
+    "total_reward_range": [
+      0.0,
+      2.0
+    ]
+  }
+}

tasks/finben_finer_ord_ner_v0/README.md

@@ -0,0 +1,52 @@
+# FinBen FINER-ORD NER Task (v0)
+
+Task module for token-level NER tagging (BIO format) on `finben_finer_ord_v0`.
+
+## Task ID
+`TA_NER_FINER_ORD_FINBEN_v0`
+
+## Dataset module
+- Dataset ID: `finben_finer_ord_v0`
+- Train: `data/finer_ord/processed/finben_finer_ord_train.jsonl`
+- Test: `data/finer_ord/processed/finben_finer_ord_test.jsonl`
+
+## Labels
+- `O`
+- `B-PER`, `I-PER`
+- `B-LOC`, `I-LOC`
+- `B-ORG`, `I-ORG`
+
+## Output schema (strict)
+Return ONLY:
+```json
+{"tags": ["O", "B-ORG", "I-ORG"]}
+```
+
+Rules:
+- Exactly one key: `tags`
+- `tags` must be a JSON array of strings
+- Length must match number of input tokens
+- Every tag must be one of the allowed BIO labels
+- No extra keys
+- No extra text
+
+<!-- REWARD_SECTION_START -->
+## Reward function (shaped, v1)
+
+This task uses a **strict-format + shaped-correctness** reward.
+
+### Total reward
+`total_reward = format_reward + correctness_reward_industry`
+
+- `format_reward` ∈ {0, 1}
+  - 1 only if output is valid JSON with exactly one key (`"tags"`), all tags valid, and length matches input tokens
+- `correctness_reward_industry` ∈ [0, 1]
+  - `0.50 * entity_span_f1`
+  - `0.40 * token_accuracy`
+  - `0.10 * exact_sequence_match`
+
+So the total reward is in **[0, 2]**.
+
+This design is stricter than free-form outputs but more informative than all-or-nothing correctness, and is better suited to RL/GRPO-style optimization.
+<!-- REWARD_SECTION_END -->
+