daloopa/financial-retrieval

Overview

This dataset contains normalized, long-form records used to benchmark multiple chatbots on financial retrieval QA. Each row represents a single (ticker, chatbot) pair answering one question. These records are derived from a verification step that extracts structured fields from each chatbot’s answer.

Columns (normalized dataset)

Column	Type	Description
ticker	string	Company identifier used for the question (e.g., AAPL, 7203:JP).
question	string	Natural-language question asked to each chatbot.
gt_answer	string	Ground-truth answer text. May contain a single value or multiple newline-separated numeric values.
gt_num	float or NaN	Parsed numeric ground truth for single-value questions. If the ground truth is a sequence, this is often NaN; use gt_answer to parse sequences.
chatbot	string (categorical)	Chatbot key: chatgpt, claude, claude_dlp, grok, gemini, perplexity.
answer	string	Raw chatbot answer text. May include rationale and/or multiple numbers.
answer_num	float or NaN	Parsed numeric prediction for single-value answers. If the answer is a sequence, this is often NaN; use answer to parse sequences.
error	int in {1..5} or NaN	Verifier-reported error category (see Error Types below). NaN indicates no error.
error_reason	string or NaN	Short rationale for the assigned error category (present only when error is set).
prec_rating	float in [1,5]	Precision rating (1–5) assigned by the verifier. Higher is better.
conc_rating	float in [1,5]	Conciseness rating (1–5). Higher is more concise.
insight_rating	float in [1,5]	Insightfulness rating (1–5). Higher reflects more helpful nuance/context.
category	string	Question category label as authored in the prompts dataset (fine-grained).
difficulty	string or int	Relative difficulty of the question (e.g., easy/medium/hard or a numeric scale, depending on source).
currency	string	Unit or currency of the ground-truth answer (e.g., USD, KRW, Percent, Units). Can have heterogeneous forms pre-normalization.
answer_link	string (URL) or NaN	Link to the chatbot’s original answer (if available in the source questions file).

Optional columns (may be present depending on generation pipeline):

• fixed_answer: string or NaN — If the verifier provided a corrected/normalized value text.
• answer_id: string or NaN — Upstream identifier linking to a specific data point or citation.
• model_version: string — Specific model or configuration version used for the run.
• timestamp: string/datetime — When the row was generated.

Error types (values of error)

1. Fiscal vs Calendar Period Confusion — Mixed fiscal vs calendar periods or used wrong reporting basis.
2. Period Shift — Picked prior/next period (e.g., Q3 vs Q2) or TTM instead of asked period.
3. Rounding / Formatting — Considerable rounding, decimal misplacement, separator/sign issues.
4. Ambiguous Interpretation — Chose an interpretation not consistent with labeling instructions.
5. Non‑Answer / Refusal — Boilerplate response or declined to answer despite available information.

Multi-value answers

Some questions have multi-value ground truths (newline-separated in gt_answer). In those cases, gt_num and answer_num may be NaN by design; parse sequences from gt_answer and answer and compare element-wise in order when computing metrics.

Recommended evaluation computations

• Exact match (error-based): treat a row as correct if error is NaN.
• Exact match (value-based): parse numeric sequences from gt_answer and answer and require exact equality element-wise when the lengths match and no element of the ground truth sequence is zero (for consistency with relative-error computations).
• Tolerance accuracy: compute relative error per element as abs(pred - gt) / abs(gt) * 100 and count the row correct if every element is within the tolerance. Skip rows with any ground-truth element equal to zero for relative-error metrics.
• Sequence length mismatch: treat as incorrect for value/tolerance metrics.

Chatbot identifiers (chatbot)

• chatgpt — ChatGPT (GPT‑5 Thinking)
• claude — Claude (Opus 4.1)
• claude_dlp — Claude (Opus 4.1) + Daloopa MCP
• grok — Grok 4
• gemini — Gemini 2.5 Pro
• perplexity — Perplexity Finance

Missing values and data quality notes

• NaN values are expected where parsing is inapplicable (e.g., multi-value answers in answer_num).
• Heterogeneous currency labels are normalized at visualization time; the raw column preserves upstream labels.
• Ratings (prec_rating, conc_rating, insight_rating) are scalar and averaged per chatbot in charts.

License and usage

Usage is intended for research and benchmarking. Please consult the repository’s LICENSE and cite the associated paper or report when using this dataset in academic work.