ProfBench: Multi-Domain Rubrics requiring Professional Knowledge to Answer and Judge
Paper • 2510.18941 • Published • 6
task_id string | domain string | prompt string | o3_response string | r1-0528_response string | grok4_response string | rubrics list |
|---|---|---|---|---|---|---|
Chem-0 | Chemistry PhD | An experimental procedure using UV-Vis for the detection of nitrogen reduction to hydrazine was determined through a method described by Watt and Chrisp. The method for hydrazine detection prepared the color reagent: p-dimethylaminobenzaldehyde, concentrated HCl, and ethanol were mixed. The chromogenic reagent was prep... | Step-1: Hydrazine concentration from the Watt–Chrisp calibration
Calibration line: A = 51 721 C + 0.1306 (C in mol L-1)
For the measured absorbance A = 0.800
C = (A – 0.1306)/51 721
= (0.800 – 0.1306)/51 721
= 0.6694 / 51 721 = 1.29 × 10⁻⁵ mol L⁻¹
Step-2: Moles of hydrazine actually produced
Total ... | To calculate the **Faradaic efficiency (FE)** and **hydrazine (N₂H₄) production rate**, we'll follow these steps:
### **Given Data:**
- **Absorbance (A)** = 0.8
- **Calibration Curve Equation:** \( y = 51721x + 0.1306 \)
- \( y \) = Absorbance
- \( x \) = Hydrazine concentration (M)
- **Total electrolyte v... | The Faradaic efficiency is calculated using the formula:
FE (%) = (n × F × moles of N₂H₄ / Q) × 100
Where:
• n = 4 (electrons required for N₂ + 4e⁻ + 4H⁺ → N₂H₄)
• F = 96485 C/mol (Faraday constant)
• moles of N₂H₄ = concentration of N₂H₄ in electrolyte × volume of electrolyte
• Q = 0.23768 C (charge passed)
The concen... | [
{
"criterion_description": "Defines the theoretical moles of hydrazine as\n\\[ n_{N_2H_4} = \\frac{n_{e^-}}{\\text{Number of electrons}} \\]",
"criterion_weight": "Minor",
"criterion_type": [
"Reasoning"
],
"o3_fulfilment": false,
"r1-0528_fulfilment": true,
"grok4_fulfilment": fal... |
Fin-0 | Finance MBA | Assume AAPL just reported its FY2018 results and has not yet reported 2019–2023 financials.
Also assume AAPL has an undrawn $250 billion revolving credit facility (RCF) that matures in 10 years with the same weighted average interest rate as AAPL’s commercial paper in FY2018 (2.18%). The RCF requires AAPL to maintain a... | ALL AMOUNTS USD millions – fiscal years end the last Saturday in September. No decimals have been rounded.
A. Baseline – FY-2018 (as reported)
• Net Revenue…………………………….. 265,595
• Interest Income……………………………. 5,475
• Interest Expense………………………….. 3,240
• Stock-Based Compensation (SBC)………. 6,700
• Deprec... | ### Financial Statement Projections for Apple Inc. (2019–2023)
Based on the 2018 Form 10-K and provided assumptions, the projections are as follows. Key metrics include revenue growth, margins, RCF activity, and cash balances. All figures are in **millions of USD**.
---
### **Income Statement**
| **Item** ... | ### Baseline (FY2018 actuals)
- Revenue: 265595
- Interest Income: 5686
- Interest Expense: 3240
- Stock-Based Compensation: 5340
- Depreciation & Amortization: 10903
- Adjusted EBITDA = Operating Income + Depreciation & Amortization + Stock-Based Compensation: 87141
- FY2018 cash and cash equivalents: 25913
- Dividen... | [
{
"criterion_description": "Calculates FY21E Net Income = (1 – Tax Rate) × [FY21E Revenues × (Gross Margin – R&D% – SG&A%) + Interest & Dividend Income – (Legacy Interest Expense + FY21E RCF Interest Expense) – Other Expense].",
"criterion_weight": "Major",
"criterion_type": [
"Reasoning"
],
... |
Cons-0 | Consulting MBA | ABC Education (the “Client”) delivers premium after-school STEM in English for ages 6-11 and is evaluating a partner-led Hong Kong entry. The entry would involve its renting rooms from international schools and using part-time Native English Teachers (NETs).
Formatting: (i) British English, (ii) No tables, (iii) Every... | Task 1 – Competitor dynamics (96 words)
Big Bang Academy
• Pricing: Cannot compute (no per-hour figure)
• Delivery: Hybrid: Kit + Video (Big Bang STEAM)
• Pedagogy: Project-Based
Blueinno
• Pricing: 10 classes – 90 min per class – HK$6,680 (HK$445 / h)
• Delivery: Learning Centre (Prince Edward innovatio... | Task 1: Competitor dynamics
Big Bang Academy: 4 classes - 60 min per class - HK$4,360. HK$1,090 per class / 1 hour = HK$1,090/h. Delivery: Hybrid: Kit + Video. Pedagogy: Project-Based.
Blueinno: 4 classes - 90 min per class - HK$3,200. (HK$3,200 / 4 classes) / 1.5 hours = HK$533.33/h. Delivery: Learning Center. Pedagog... | ### Task 1: Competitor dynamics
(i) Big Bang Academy:
1. Pricing: Cannot compute (no per-hour figure).
2. Delivery: Hybrid: Kit + Video.
3. Pedagogy: Project-Based.
(ii) Blueinno:
1. Pricing: 10 classes - 90 min per class - HK$6,680 (HK$445/h).
2. Delivery: Learning Center.
3. Pedagogy: Project-Based.
... | [
{
"criterion_description": "Calculates Big Bang Academy's hourly rate by taking the total cost / the total class hours across all classes.",
"criterion_weight": "Critical",
"criterion_type": [
"Reasoning"
],
"o3_fulfilment": false,
"r1-0528_fulfilment": false,
"grok4_fulfilment": f... |
Phys-0 | Physics PhD | Semiclassical self-gravity is sometimes proposed as an alternative to a quantized theory of gravity, mainly as a model to test for and rule out experimentally, in order to find evidence for the necessity to quantize gravity. In this model, the spatial wave function of quantum matter nonlinearly reacts to its own gravit... | We have to evaluate the relative phase that builds up between the two Stern–Gerlach branches |a〉 and |b〉 of the 100-atom silicon cluster while the two packets fly freely for
t = τ = 10 ns = 1.0 × 10-8 s
and remain separated by a distance d that is “much larger than the particle radius”.
Because d ≫ R the mutua... | ### Step 1. Understand the Setup
A particle composed of 100 silicon atoms (total spin \( \frac{1}{2} \)) is placed in a Stern–Gerlach apparatus. Its wave function is split into two branches:
\[
\psi = \sqrt{\frac{1}{3}} |a\rangle + \sqrt{\frac{2}{3}} |b\rangle
\]
The branches follow different trajectories, separated... | The semiclassical model sources the gravitational potential with the position-space probability density ρ(r) = m |ψ(r)|², where ψ is the total wavefunction and m is the total mass of the particle.
The wavefunction is ψ = √(1/3) ψ_a + √(2/3) ψ_b, with ψ_a and ψ_b localized at separated positions, no overlap since d >> ... | [
{
"criterion_description": "Analyzes that a sphere of 100 silicon atoms has a mass of $2800 - 2810 \\ \\text{u}$ or $4.6-4.7 \\ 10^{-24}\\,\\text{ kg}$.",
"criterion_weight": "Minor",
"criterion_type": [
"Reasoning"
],
"o3_fulfilment": true,
"r1-0528_fulfilment": true,
"grok4_fulfi... |
Chem-1 | Chemistry PhD | A master’s student who recently joined our group is eager to challenge himself by synthesizing compounds featuring Al–Al and Ga–Ga multiple bonds, inspired by recent works in the field. Some researchers recently reported the successful synthesis of a compound with a B–B triple bond, and the student aims to prepare the ... | Internal note
“Mapping the synthetic space for E–E triple bonds (E = B, Al, Ga)”
The aim is to move beyond the preparation of two isolated compounds and to obtain comparative data that reveal how the difficulty of forging an E≡E bond changes on descending group 13. The outline below follows the order requested by t... | Of course. Here is a detailed note designed to guide the master's student in developing a systematic and insightful research plan.
***
**To:** [Master's Student]
**From:** [PI/Advisor]
**Date:** [Current Date]
**Subject:** Research Plan: Investigating the Feasibility of Group 13 Triple Bonds (Al–Al, Ga–Ga)
Welcome t... | Below is a structured note prepared for the master's student, based on the query and drawing from the provided literature (e.g., the 2012 Science paper on diboryne synthesis and the RSC paper on B-B triple bond characterization via 11B-11B J-coupling). This note aims to guide a systematic experimental plan for explorin... | [
{
"criterion_description": "Describe that the ^13C NMR signal for the carbon directly bonded to the group-13 center generally shows a noticeable upfield shift relative to the single-bond analogue, consistent with increased bond order and altered electronic environment; precise shift size is system-dependent.",
... |
Fin-1 | Finance MBA | ABC Capital Partners is facing a liquidity crisis in one of its funds and is seeking to sell a large portfolio of assets (the "Portfolio") in order to free up liquidity. The Portfolio has been on the market for six months, but given the distressed nature of the assets in the Portfolio, the process has not yet yielded s... | NOTE All dollar figures in millions and rounded to the nearest $1 mm unless a “$ /sh” figure (1 decimal).
Percentages are rounded to 0.1%.
INPUT SUMMARY
• Q1’25 quarterly NOI = $32 mm
• Investment-property book value = $2,469 mm
• Total assets = $2,574 mm
• Mortgages (ex-held-for-sale) = $590 mm
• Credit f... | 1. REIT's Current Metrics
Source: Nexus REIT Q1'25 Financials
(Assumed figures since the actual PDF is not accessible, but we'll proceed with typical REIT metrics.)
Total Assets: $2,000mm
Total Debt: $800mm
Quarterly NOI: $25mm → Annualized NOI = $100mm
Book Value (NAV): $1,200mm
Shares Outstanding: 100mm
NAV pe... | ### 1. REIT's current cap rate, debt-to-total-assets, and share price premium/discount to NAV
To determine these, I first need the REIT's key financials as of Q1 2025 from the fetched data.
From the financial statements and MD&A:
- Investment properties fair value: $2,469 million
- Quarterly NOI: $32 million
- Annual... | [
{
"criterion_description": "States that the REIT's book value is $2,574mm",
"criterion_weight": "Minor",
"criterion_type": [
"Extraction (recall)"
],
"o3_fulfilment": true,
"r1-0528_fulfilment": false,
"grok4_fulfilment": false
},
{
"criterion_description": "States that the... |
Cons-1 | Consulting MBA | The year is 2017 and our consulting firm's Automotive Manufacturing Excellence team has been brought in to advise Tesla Motors. In the face of overwhelming demand for its newly-revealed Model 3, the company is looking to rapidly scale past its "production hell" to achieve operational and financial leverage. The Head of... | 1. OEE ― VARIABLES AND FORMULAS
• Availability = Operating Time ÷ Planned-Production Time
– Planned-Production Time = Shift Length – Planned Downtime
– Operating Time = Planned-Production Time – (Unplanned Downtime + Setup/Changeover)
• Performance = (Ideal Cycle Time ... | ### **Report to: Head of Operations, Tesla Motors**
### **From: Automotive Manufacturing Excellence Team**
### **Date: October 26, 2017**
### **Subject: OEE Analysis & Improvement Roadmap for Model 3 Production**
### **1. OEE Explanation**
Overall Equipment Effectiveness (OEE) is the gold standard metric for measurin... | 1. OEE explanation: Overall Equipment Effectiveness (OEE) measures how effectively a manufacturing process is utilized compared to its full potential, calculated as the product of three variables: Availability × Performance × Quality. Availability is Run Time divided by Planned Production Time, where Planned Production... | [
{
"criterion_description": "Calculates baseline Availability utilizing formula: Availability = Run Time / Planned Production Time",
"criterion_weight": "Major",
"criterion_type": [
"Reasoning"
],
"o3_fulfilment": true,
"r1-0528_fulfilment": true,
"grok4_fulfilment": true
},
{
... |
Phys-1 | Physics PhD | In thin film layer, we usually have anisotropy, like PMA (perpendicular magnetic anisotropy), etc., due to which we have an anisotropy in the direction of the magnetic moment. Magnetic moments tend to align along the direction of PMA or low energy field. PMA or other anisotropies are generally fields, and their energy ... | Given data
• Number density of magnetic ions: n = 100 ions m⁻³
• Magnetic moment per ion: µ = 3 × 10⁻²⁴ A m²
• Perpendicular anisotropy field: µ₀H_k = 0.5 T
• Film thickness: t = 10 mm = 0.01 m
• Intercept of the (E_eff · t)–t plot is 0 mJ m⁻² ⇒ no interfacial term, so the measured energy density is the... | ### Step 1. Understand the Anisotropy Energy Density
In magnetic materials with uniaxial anisotropy (like PMA), the anisotropy energy density is given by:
\[
K = \frac{1}{2} \mu_0 M_s H_k
\]
where:
- \( \mu_0 = 4\pi \times 10^{-7} \text{H/m} \) is the permeability of free space,
- \( M_s \) is the saturation magnet... | First, note that the unit for the density of magnetic ions is likely a typo and meant to be 100 ions/mm³ instead of m³, as the given value with m³ leads to unrealistically low magnetization for a bulk material. With 1 mm³ = 10^{-9} m³, the number density n = 100 / 10^{-9} = 10^{11} ions/m³.
The saturation magnetizatio... | [
{
"criterion_description": "Computes the saturation magnetization using the equation $M_{\\text{sat}} = \\mu_{\\text{ion}} N_{\\text{mag}}$, where $\\mu_{\\text{ion}} = 3 \\times 10^{-24} \\, \\text{A}\\cdot\\text{m}^2$ is the magnetic moment per ion, and $N_{\\text{mag}} = 1.0 \\times 10^{2} \\, \\text{ions/m}... |
Chem-2 | Chemistry PhD | "Currently, producing life-saving medicines through organic synthesis is crucial in the pharmaceutic(...TRUNCATED) | "1. Preferred IUPAC name of product X \n4-Bromo-2-chloropyridine\n\n2. ¹H-NMR multiplicities (thre(...TRUNCATED) | "### 1. Preferred IUPAC Name of Product X \nThe product X is **3-chloro-4-bromopyridine-2-d** (or e(...TRUNCATED) | "### 1. Preferred IUPAC Name of Product X\n\n3-bromo-2-chloro-4-deuteriopyridine\n\n### 2. Signal Mu(...TRUNCATED) | [{"criterion_description":"Provides the final signal for the product is (Z, Z, Z, d, d)","criterion_(...TRUNCATED) |
Fin-2 | Finance MBA | "Circle Internet Group is a blockchain-focused financial services and payments company founded in 20(...TRUNCATED) | "Step 1 – Identify the relevant numbers \n• Offer price = $31.00 per share \n• Base deal si(...TRUNCATED) | "Based on the provided information, Circle Internet Group had a pre-IPO valuation of $6.80 billion a(...TRUNCATED) | "1. The total dollar value of proceeds from the IPO before underwriting fees and the exercise of gre(...TRUNCATED) | [{"criterion_description":"Computes the portion of gross IPO proceeds retained by selling stockholde(...TRUNCATED) |
End of preview. Expand in Data Studio
nvidia/ProfBench（複製）
データの内容
人手でアノテーションされた 40 件の専門的な課題と、それに対する 3 モデルの応答、評価基準のデータ。1 行は 1 課題で、prompt に課題文、rubrics に 15〜59 個の評価基準、domain に領域、o3_response などに各モデルの応答が入る。領域は Chemistry PhD、Physics PhD、Finance MBA、Consulting MBA の 4 種類で、一部の応答は Grok を使って作られており、カードは LLM の評価に使うことを想定している。
データの概要
| 項目 | 値 |
|---|---|
| 行数 | 40（1 構成・1 分割の合計） |
| 列数 | 7 |
| 容量 | 1.20 MB（ファイル 4 本） |
| ファイルの形式 | pdf 1、jsonl 1 |
| 言語 | en |
| 公開 | public |
構成と分割
| 構成（config） | 分割（split）ごとの行数 | 列数 |
|---|---|---|
default |
test 40 |
7 |
列（構成 default・分割 test）
| 列 | 型 | 先頭の行の値 |
|---|---|---|
task_id |
string |
Chem-0 |
domain |
string |
Chemistry PhD |
prompt |
string |
An experimental procedure using UV-Vis for the d… |
o3_response |
string |
Step-1: Hydrazine concentration from the Watt–Ch… |
r1-0528_response |
string |
To calculate the **Faradaic efficiency (FE)** an… |
grok4_response |
string |
The Faradaic efficiency is calculated using the … |
rubrics |
list<struct{criterion_description: string, criterion_weight: string, criterion_type: list<string>, o3_fulfilment: bool, r1-0528_fulfilment: bool, …}> |
[23 件] |
行数と列は HF のデータセットビューアの集計（元のデータセットの現在の版で数えたもの）。全データセットの一覧は HYUKI5/dataset-catalog にある。
複製について
nvidia/ProfBench をリビジョン f5c471cf2f277c264621e4b5c0cad5bb466f9260 の時点でそのまま複製したもの。中身は変えていない。判断モデル Laya（convaiinnovations/laya）の微調整に使う候補として、調査メモ（Laya 学習用データセット調査）に載せたデータを集めている。
出典
- 元のデータセット: https://huggingface.co/datasets/nvidia/ProfBench
- 複製したリビジョン:
f5c471cf2f277c264621e4b5c0cad5bb466f9260（2026-03-04 06:27:42+00:00） - 元のライセンス表記:
other
論文
アノテーションの方法
データセットの作成日:
9/24/2025
微調整に使うときの注意
- 調査メモの台帳での区分: B（継承条件・gated・モデル由来の条件・上流の規約などが付く）。区分は記載の文言で機械的に付けたもので、法的な判定ではない。学習に入れる前に元のカードとライセンスの原文で確かめる。
- Laya の入力は 2,048 token まで（ModernBERT-large / mmBERT）。長い本文は先頭から黙って切られるので、切れる行は事前に除くか、要約・切り出しをしてから入れる。
- 分割は記事・文書単位で行い、同じ本文から作った別の問いを学習側と較正側に分けない。温度と閾値は学習に使っていない分割で別々に当てる。
調査メモの記述
| メモ | 節 | ライセンス（記載） | 確認 | 注記 |
|---|---|---|---|---|
nvidia_nemotron.md |
Q2. 金融・経済・ビジネス・ニュース・法務に関係するサブセットはあるか。ラベルはあるか (L79) | nvidia-evaluation-dataset-license（評価用。学習への利用可否は条文未読で未確認） | 確認済 | 原文注記: 40 件のみ。評価専用の可能性が高く、学習に入れるのは避ける |
元のカード（日本語訳）
以下は元のデータセットのカードの本文を日本語に訳したもの（機械翻訳）。訳に疑問があるときは、末尾の原文を確かめる。
データセットの説明:
Leaderboard | ブログ | 論文 | データ | コード | Nemo Evaluator SDK
40 件の人手でアノテーションされた課題にまたがる、3000 組を超えるルーブリックと応答の組。専門的な課題に答えたレポートを扱い、対象領域は博士課程レベルの STEM（化学、物理）と専門サービス（金融サービス、経営コンサルティング）である。
このデータセットは商用・非商用のどちらにも使える。
データセットの所有者:
NVIDIA Corporation
データセットの作成日:
9/24/2025
ライセンス/利用規約:
NVIDIA Evaluation Dataset License
想定される用途:
専門的な課題で LLM を評価したい研究者と開発者。ProfBench は、数十のベンチマークを横断する評価のための統一インターフェースを備えた Nemo Evaluator SDK の一部として使うことを推奨する。
データセットの特徴づけ:
** データ収集の方法
- [ハイブリッド: 人手・合成・自動]
** ラベル付けの方法
- [人手]
データセットの形式:
テキスト。
データセットの量:
40 件のレコード
各レコードには次のフィールドが含まれる:
- ID: 各サンプルの一意の識別子
- Domain: 領域。Chemistry PhD / Physics PhD / Finance MBA / Consulting MBA のいずれか
- Prompt: 大規模言語モデル（LLM）への指示
- Rubrics: 最終的なモデル出力を評価するために使う、15〜59 個の一意の基準
- Model Responses: OpenAI o3 / xAI Grok4 / DeepSeek R1-0528 による 3 件の応答
このデータセットの一部は Grok を使って作成された。
総容量: 1 MB。
倫理的な配慮:
NVIDIA は、信頼できる AI は共通の責任であると考えており、広範な AI アプリケーションの開発を可能にするための方針と慣行を整えている。利用規約に従ってダウンロードまたは使用する場合、開発者は社内のモデルチームと協力して、このモデルが関連する業界と用途の要件を満たし、想定外の製品の誤用に対処することを確認すべきである。 モデルの品質、リスク、セキュリティ上の脆弱性、または NVIDIA の AI に関する懸念は、こちら から報告すること。
引用:
ProfBench が役に立った場合は、以下を引用することを検討すること:
@misc{wang2025profbenchmultidomainrubricsrequiring,
title={ProfBench: Multi-Domain Rubrics requiring Professional Knowledge to Answer and Judge},
author={Zhilin Wang and Jaehun Jung and Ximing Lu and Shizhe Diao and Ellie Evans and Jiaqi Zeng and Pavlo Molchanov and Yejin Choi and Jan Kautz and Yi Dong},
year={2025},
eprint={2510.18941},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2510.18941},
}
原文（英語）
Dataset Description:
Leaderboard | Blog | Paper | Data | Code | Nemo Evaluator SDK
More than 3000 rubric-response pairs across 40 human-annotated tasks presenting reports addressing professional tasks across PhD STEM (Chemistry, Physics) and Professional Services (Financial Services, Management Consulting) domains.
This dataset is ready for commercial/non-commercial use.
Dataset Owner(s):
NVIDIA Corporation
Dataset Creation Date:
9/24/2025
License/Terms of Use:
NVIDIA Evaluation Dataset License
Intended Usage:
Researchers and developers seeking to evaluate LLMs on Professional Tasks. We recommend use of ProfBench as part of Nemo Evaluator SDK, which supports a unified interface for evaluation across tens of benchmarks.
Dataset Characterization:
** Data Collection Method
- [Hybrid: Human, Synthetic, Automated]
** Labeling Method
- [Human]
Dataset Format:
Text.
Dataset Quantification:
40 records
Each record contains the following fields:
- ID: Unique identifier for each sample
- Domain: Chemistry PhD / Physics PhD / Finance MBA / Consulting MBA
- Prompt: Instruction for the Large Language Model (LLM)
- Rubrics: 15-59 unique criterion used to assess the final model output
- Model Responses: 3 responses from OpenAI o3 / xAI Grok4 / DeepSeek R1-0528
Some portions of this dataset were created with Grok.
Total Storage: 1 MB.
Ethical Considerations:
NVIDIA believes Trustworthy AI is a shared responsibility and we have established policies and practices to enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.
Please report model quality, risk, security vulnerabilities or NVIDIA AI Concerns here.
Citation:
If you found ProfBench helpful, please consider citing the below:
@misc{wang2025profbenchmultidomainrubricsrequiring,
title={ProfBench: Multi-Domain Rubrics requiring Professional Knowledge to Answer and Judge},
author={Zhilin Wang and Jaehun Jung and Ximing Lu and Shizhe Diao and Ellie Evans and Jiaqi Zeng and Pavlo Molchanov and Yejin Choi and Jan Kautz and Yi Dong},
year={2025},
eprint={2510.18941},
archivePrefix={arXiv},
primaryClass={cs.CL},
url={https://arxiv.org/abs/2510.18941},
}
- Downloads last month
- 53