HenryJ12/InvestorLastExam

Investor's Last Exam v1

86 financial research and forecasting tasks, with 1,472 English rubrics, the solver and judge system prompts, and complete function-tool definitions. There are 41 structured tasks, 37 report tasks, and 8 market-forecast tasks.

Load

from datasets import load_dataset

tasks = load_dataset("HenryJ12/InvestorLastExam", split="test")
rubrics = load_dataset("HenryJ12/InvestorLastExam", "rubrics", split="test")

For local files:

tasks = load_dataset("parquet", data_files={"test": "data/test.parquet"}, split="test")
# The equivalent JSONL is also included:
import json
from datasets import Dataset
with open("data/test.jsonl", encoding="utf-8") as stream:
    tasks_json = Dataset.from_list([json.loads(line) for line in stream])

The JSONL example above preserves ISO-8601 offsets as literal strings, without automatic datetime inference. The default Parquet configuration does the same.

Upload this directory's contents to a Hugging Face dataset repository:

hf upload HenryJ12/InvestorLastExam . --repo-type dataset

Fields

• id, task_id: the same stable public task code, such as O37 or N16.
• company, ticker, category_id, category_name, cutoff: task metadata.
• question: complete English public solver instruction.
• submission_mode: structured, report, or market.
• solver_system_prompt: the common system instruction, verified against all 86 recorded solver requests.
• judge_kind, judge_system_prompt: the actual grading branch and its complete system instruction. Report instructions include the question-specific rubric IDs and output-schema instructions. Market tasks have a null judge prompt.
• rubrics: typed rubric records, including original weights, normalized max_score, exact fractional maxima, and the complete definition_json.
• tools_json: all 11 solver function definitions for this task, including its exact submit_answer schema. Parse with json.loads.
• submission_schema_json: the task-specific public submission schema.
• judge_tools_json: the structured judge's submission function, or an empty list for report and market grading. Judges have no retrieval/research tools.
• judge_input_template_json: the judge-input structure and reference material, with explicit submission placeholders.
• judge_response_schema_json: judge output schema; report answer-span IDs are populated from the submitted answer at evaluation time.
• ground_truth_json, evaluation_policy_json: frozen evaluation truth and score-allocation/formula parameters. These are evaluator inputs.
• source_sha256: SHA-256 of the nested public task record, excluding this field, serialized as UTF-8 JSON with sorted keys, no ASCII escaping, and compact separators.

The *_json fields preserve heterogeneous schemas and complete rubric details without Arrow type conflicts. tasks/*.json provides the same records with ordinary nested JSON objects for users who prefer that representation.

Rubric types

type has exactly two values:

• llm_judge: evaluated by the semantic/report judge (1,242 items).
• programmatic: executed by the deterministic scoring code (230 items).

This is the actual evaluation route, not a relabeling of the author's original verification field. In free-text report tasks, numerical criteria are evaluated by the report judge and therefore remain llm_judge, even if their underlying rule is numeric. Structured-task deterministic/future-truth rules, the seven numerical/probability rules in TP05, and all eight market losses are programmatic. The verification tag, current criteria, formulas, thresholds and source evidence are included in definition_json. Use max_score for normalized points; original author weights are not always on a 100-point scale.

Prompts and tools

• prompts/solver.txt: the shared solver system prompt.
• prompts/judge_structured.txt: complete structured-judge system prompt.
• prompts/judge/<id>.txt: complete system prompt for every LLM-graded task.
• prompts/judge_strict_supplement.txt: the strictness supplement already included in the complete prompts, provided separately for inspection.
• prompts/judge_index.json: per-task grading routes, tools and schema templates.
• prompts/answer_spans.py: the original, pure report-span segmentation function.
• schemas/answer_spans.json: the corresponding span-record schema; offsets are Python string offsets (Unicode code points), not UTF-8 byte offsets.
• tools/solver_toolsets.json: deduplicated common tool definitions.
• tools/submit_answer.json: the public submission-tool definition for each task.
• tools/judge_structured.json: the judge submission function's base definition.

The solver tools are search, read, download, bash, tables, table_lookup, table_query, table_read, table_cancel, table_download, and submit_answer. The structured judge uses submit_semantic_grade_v11; report judges return schema-constrained JSON and do not use tools.

For structured judge inputs, replace anonymous_answer with the submitted JSON object. For report inputs, replace submitted_answer with the complete report and answer_spans with the immutable answer-span records. TP05's structured payload is serialized as JSON for its semantic items; its numerical items remain scored by their original rules. Astra and OSS use the same grading instructions; the provider protocol differs. An objective market task makes no judge-model call.

Give the solver only its question, solver system prompt and public tools/schema; rubrics, judge inputs and future truth belong to the evaluator. Tool definitions describe the ProphetBenchEnv interface; live data/tool backends are not included in this dataset package.

Files

data/test.parquet and data/test.jsonl contain the same 86 tasks. data/rubrics.parquet and data/rubrics.jsonl contain the same 1,472 rubric rows. manifest.json records counts and exported-file checksums.