Update README.md

README.md

@@ -1,3 +1,167 @@
 ---
-license: apache-2.0
+language:
+- zh
+- en
+task_categories:
+- question-answering
+tags:
+- finance
+- financial-research
+- information-retrieval
+- web-search
+- agents
+- benchmark
+pretty_name: FinFIRST
+size_categories:
+- n<1K
+configs:
+- config_name: default
+  data_files:
+  - split: test
+    path: data/test.jsonl
 ---
+
+# FinFIRST: Financial Information Retrieval, Sourcing and Traceability
+
+**FinFIRST** is a benchmark for evaluating the retrieval quality, source reliability, and answer traceability of financial search agents. It was jointly developed by Ant Group and China International Capital Corporation Limited (CICC).
+
+Unlike benchmarks based only on final-answer matching, FinFIRST uses atomic rubrics to assess whether an agent retrieves the correct information, verifies authoritative and time-valid sources, applies consistent definitions, and produces a reproducible answer.
+
+## Highlights
+
+- **123** expert-authored and verified financial research tasks.
+- **50+** finance professionals involved in construction and quality control.
+- **701** atomic criteria with a combined rubric weight of **12,300** points.
+- An **18-field taxonomy**, a **six-axis coverage blueprint**, and **138 source categories**.
+- A six-stage review process with a **9.78% acceptance rate**.
+- **15 model configurations** evaluated under a unified agent and tool environment.
+
+![Capability comparison between FinFIRST and representative financial benchmarks](assets/benchmark-comparison.png)
+
+## What FinFIRST Evaluates
+
+Each task is decomposed into independently assessable criteria across three dimensions:
+
+- **Raw-information acquisition:** retrieving the correct facts and values for the required entity, period, unit, definition, and data version.
+- **Source verification:** using authoritative, relevant, and time-valid sources.
+- **Computation and answer formation:** performing calculations and synthesis correctly and supporting the conclusion with evidence.
+
+This design distinguishes a fully supported answer from one that is correct only by chance and attributes failures to retrieval, sourcing, definition alignment, or computation.
+
+## Dataset Construction
+
+FinFIRST was built through three stages: scenario-driven data preprocessing, expert task construction, and quality control.
+
+![FinFIRST dataset construction and quality-control pipeline](assets/data-construction-pipeline.png)
+
+For every task, experts assembled a reference package containing the answer, supporting sources, key data, definitions and versions, calculations, valid alternatives, and numerical tolerances. Candidate tasks then passed through value and scope review, independent re-solving, cross-validation, rubric audit, model stress testing, and consistency checking. All revisions and inclusion decisions were made by financial experts.
+
+## Dataset Statistics
+
+| Dimension | Value |
+|---|---:|
+| Total tasks | 123 |
+| Chinese / English | 74 (60.2%) / 49 (39.8%) |
+| Multiple related subquestions | More than 80% |
+| Explicit computation required | 61.8% |
+| Multiple sources required | 32.5% |
+| No prescribed source | 75.6% |
+| Publicly accessible sources | 100% |
+| Atomic criteria | 701 |
+
+The tasks cover Mainland China, the United States, Hong Kong, and other markets, with research objects spanning companies, industries, regional economies, and financial instruments.
+
+![FinFIRST dataset distribution](assets/dataset-distribution.png)
+
+## Data and Usage
+
+The repository contains a single `test` split with 123 JSONL records:
+
+```text
+README.md
+assets/*.png
+data/test.jsonl
+```
+
+```python
+from datasets import load_dataset
+
+dataset = load_dataset("inclusionAI/FinFIRST", split="test")
+
+print(dataset[0]["query"])
+print(dataset[0]["answer"])
+print(dataset[0]["rubric"])
+```
+
+### Schema
+
+Each record contains 24 string-or-null fields:
+
+- **Core:** `original_id`, `id`, `query`, `answer`, `rubric`, `rubric_annotated`
+- **Task structure:** `question_decomposition`, `relation_type`, `conditional_filtering`
+- **Research target:** `subject_category`, `metric`, `measurement_scope`, `region`
+- **Time and verification:** `time_constraint_type`, `time_description`, `verification_method`
+- **Output requirements:** `calculation_type`, `output_format_requirement`, `language`
+- **Sources and intent:** `source_specification`, `source_accessibility`, `source_mainstreamness`, `source_count`, `intent_category`
+
+`answer` is always stored as a string. Excel percentages were converted to their requested presentation format, such as `0.0083` to `0.83%`. Missing cells are represented as `null`, while the literal category value `None` is preserved.
+
+### Atomic Rubric Example
+
+The example below requires an agent to combine Meta's official disclosures with an IAB/PwC report. The final answer is **72.45%**, but the rubric separately checks source identification, data extraction, aggregation, calculation, precision, and units.
+
+![Example of a FinFIRST task and atomic rubric](assets/rubric-example.png)
+
+## Evaluation
+
+All baseline models use the same ReAct-style agent framework with **Web Search**, **Visit**, and **Python** tools. GLM-5.1 serves as the automated judge. On 50 randomly sampled instances adjudicated by eight financial experts, agreement between the automated judge and human labels reached a Cohen's $\kappa$ of **0.816** at the atomic-criterion level.
+
+FinFIRST reports four complementary metrics:
+
+- **Atomic:** unweighted pass rate over all atomic criteria.
+- **Loose Pass:** partial-credit score using expert-assigned criterion weights.
+- **Strict Pass:** percentage of tasks for which every atomic criterion passes.
+- **Unsupported-Correct Rate (UCR):** share of correct answers with incomplete evidence.
+
+$$
+\mathrm{UCR}=\frac{N(\text{correct, partial evidence})}
+{N(\text{correct, partial evidence})+N(\text{correct, fully traceable})}.
+$$
+
+Lower UCR indicates that correct answers are more consistently accompanied by complete and verifiable evidence.
+
+## Results
+
+### Overall Performance
+
+![Overall FinFIRST benchmark results](assets/overall-results.png)
+
+GPT-5.6-Sol achieves the highest Strict Pass score at **71.54%**. Claude-Opus-5 reaches **87.61%** Loose Pass but only **69.11%** Strict Pass, showing that strong partial performance can still hide incomplete task execution. Across all evaluated models, computation and answer formation remain weaker than raw-information acquisition.
+
+### Unsupported-Correct Rate
+
+![Unsupported-Correct Rate results](assets/unsupported-correct-rate.png)
+
+Among 1,150 correct-answer instances, 201 do not satisfy the complete evidence requirements, producing a micro-average UCR of **17.48%**. Among open-source models, Ling-3.0-Flash-Fin achieves the lowest UCR and the highest source-verification score.
+
+### Slice Analysis
+
+![FinFIRST slice analysis](assets/slice-analysis.png)
+
+Performance declines from easy to hard tasks, multi-source tasks are consistently more difficult than single-source tasks, and model performance varies across languages and markets.
+
+## Intended Use and Limitations
+
+FinFIRST is intended for evaluating financial search and deep-research agents, studying evidence-grounded retrieval and numerical reasoning, developing rubric-based judges, and analyzing model failures.
+
+Some tasks depend on information available before a specified cutoff date and may become outdated after later disclosures or revisions. The dataset includes complete answers and rubrics, so it should not be treated as a hidden-label leaderboard set, and training on it may contaminate evaluation. FinFIRST is a research benchmark and does not constitute investment advice.
+
+## License
+
+No open-source license is currently declared. Before making the repository public, select a license consistent with the ownership of the dataset and the usage terms of any third-party material.
+
+## Paper
+
+**arXiv:** _Link to be added._
+
+<!-- Replace the line above with the arXiv URL when the paper is available. -->