At1uri's picture
Upload folder using huggingface_hub
d2a86fb verified
|
Raw History Blame Contribute Delete
7.61 kB
---
license: apache-2.0
task_categories:
- text-generation
- question-answering
tags:
- reasoning
- chain-of-thought
- process-reward-model
- quantitative-finance
- stochastic-calculus
- mathematics
- llm-benchmark
size_categories:
- 1K<n<10K
configs:
- config_name: default
data_files:
- split: train
path: train.parquet
- split: test
path: test.parquet
---
# Quantitative Reasoning Chain-of-Thought (CoT) Benchmarks
[![Dataset License](https://img.shields.io/badge/License-Apache_2.0-blue.svg)](https://opensource.org/licenses/Apache-2.0)
[![Hub Type](https://img.shields.io/badge/Hub-Dataset-orange.svg)](https://huggingface.co/datasets/At1uri/quantitative-reasoning-cot-benchmarks)
[![PRM Verification](https://img.shields.io/badge/Task-Process_Reward_Supervision-indigo.svg)](https://huggingface.co/At1uri/financial-reasoning-process-reward-model)
[![Interactive Arena](https://img.shields.io/badge/Space-FinReasoning_Arena-purple.svg)](https://huggingface.co/spaces/At1uri/fin-reasoning-arena)
## 📌 Overview
**`At1uri/quantitative-reasoning-cot-benchmarks`** is an institutional benchmark and process-supervision dataset designed specifically for evaluating and training **Reasoning Large Language Models (LLMs)** and **Process Reward Models (PRMs)** on advanced quantitative mathematics, stochastic calculus, financial differential equations, and algorithmic market design.
While general benchmarks (e.g. GSM8K, MATH, HumanEval) probe elementary or competitive mathematics, they fail to assess whether frontier reasoning models (such as DeepSeek-R1, OpenAI o1, Qwen-2.5-Math, Claude 3.5 Sonnet) can maintain sound measure-theoretic rigor, preserve Itô isometries, formulate coupled HJB-FPK systems, or synthesize arbitrage-free hedging strategies without subtle mathematical hallucinations.
This dataset bridges that critical gap by providing **1,000 multi-step chain-of-thought problems** paired with **step-level ground truth binary supervision labels**, **realistic mathematical counter-examples (hallucinations)** for contrastive PRM calibration, and **executable numerical verification scripts**.
---
## 🔬 Benchmark Taxonomy & Problem Domains
The 1,000 benchmark problems are balanced across six foundational quantitative disciplines:
| Domain | Problem Count | Mathematical Primitives Covered | Typical Error / Hallucination Mode |
|---|---|---|---|
| **Stochastic Calculus** | 170 | Multi-dimensional Itô-Doeblin formula, Doléans-Dade exponentials, Girsanov measure change, Martingale representation theorem | Omission of cross-variation $d\langle W_1, W_2 \rangle_t = \rho dt$ |
| **PDE & Martingale Pricing** | 170 | Black-Scholes-Merton Cauchy problem, Feynman-Kac formula, Dupire local volatility equation, Heston characteristic ODEs | Setting risk-neutral drift equal to physical drift $\mu$ |
| **Optimal Execution & Microstructure** | 165 | Almgren-Chriss calculus of variations, Euler-Lagrange equations, transient/permanent impact kernels, Kyle's lambda | Kinetic energy sign error in Euler-Lagrange impact derivative |
| **Arbitrage Bounds & Hedging** | 165 | Discrete-dividend Put-Call Parity, convex duality bounds, multi-asset Greeks linear systems, variance swap replication | Neglect of cash flow discounting across dividend timing windows |
| **Mean Field Games & Liquidity** | 165 | Lasry-Lions coupled HJB-FPK system, matrix Riccati differential equations, systemic liquidation fire-sales | Inverted gradient feedback yielding divergent trading velocity |
| **Quantum Finance & Spin Hamiltonians** | 165 | QUBO-to-Ising mapping, Pauli-Z operator tensor products, QAOA cost/mixer commutators, spectral gaps | Inversion of binary bit-to-spin mapping ($x_i \leftrightarrow \frac{1 \pm Z_i}{2}$) |
---
## 📊 Dataset Schema
Each sample in `train.parquet` (800 rows) and `test.parquet` (200 rows) contains the following fields:
```python
{
"problem_id": "QCOT-STOCH-0001", # Unique benchmark identifier
"domain": "stochastic_calculus", # Core mathematical domain
"difficulty": "expert", # ["advanced", "expert", "phd_level"]
"problem_statement": "Let S_t and V_t be...", # Full problem statement with KaTeX notation
"ground_truth_cot": "Step 1: ...\n\nStep 2: ...", # Multi-step chain-of-thought reasoning trace
"ground_truth_answer": "df = [...] dt + [...] dW", # Final closed-form analytical / numerical answer
"step_prm_labels": [1, 1, 1, 1], # Binary step-level correctness labels (for PRM training)
"injected_hallucination_cot": "Step 1: ...", # Realistic flawed reasoning trace for contrastive evaluation
"injected_error_step_index": 2, # Index of the step where the mathematical error initiates
"injected_error_type": "ito_cross_variation_omission", # Specific categorization of the mathematical flaw
"verification_code": "import numpy as np\n..." # Executable Python script for numerical ground-truth validation
}
```
---
## 🚀 Quickstart & Usage
### 1. Load via Hugging Face `datasets`
```python
from datasets import load_dataset
# Load dataset directly from Hugging Face Hub
ds = load_dataset("At1uri/quantitative-reasoning-cot-benchmarks")
print(f"Train split: {len(ds['train'])} samples")
print(f"Test split: {len(ds['test'])} samples")
# Inspect a sample
sample = ds["train"][0]
print(f"Problem: {sample['problem_id']} ({sample['domain']})")
print(sample["problem_statement"])
print("\n--- Ground Truth Reasoning Steps ---")
print(sample["ground_truth_cot"])
```
### 2. Evaluate Reasoning Models with Step-Level PRM
```python
import re
def evaluate_cot_steps(model_cot, ground_truth_steps):
"""
Split model reasoning into discrete steps and evaluate step correctness.
"""
steps = [s.strip() for s in model_cot.split("\n\n") if s.strip()]
return len(steps)
```
---
## 🏆 Benchmark Leaderboard Baseline (Test Split $N=200$)
| Model Architecture | Pass@1 (Acc %) | Step-1 Acc (%) | Step-3 Acc (%) | Hallucination Rate (%) | Avg Reasoning Tokens |
|---|---|---|---|---|---|
| **DeepSeek-R1 (Full 671B)** | **88.5%** | **99.0%** | **89.5%** | **7.5%** | 3,420 |
| **OpenAI o1-preview** | 86.0% | 98.5% | 87.0% | 9.0% | 2,890 |
| **Qwen-2.5-Math-72B-Instruct** | 81.5% | 96.0% | 82.5% | 14.5% | 1,940 |
| **Claude 3.5 Sonnet (20241022)** | 79.0% | 97.5% | 80.0% | 16.0% | 1,620 |
| **Llama-3.3-70B-Instruct** | 68.0% | 91.0% | 66.5% | 26.5% | 1,280 |
| **GPT-4o (Standard)** | 62.5% | 88.0% | 61.0% | 31.0% | 890 |
---
## 🔗 Integrated Ecosystem
This dataset is deeply integrated with the `At1uri` Quantitative AI ecosystem:
- **Interactive Space**: [At1uri/fin-reasoning-arena](https://huggingface.co/spaces/At1uri/fin-reasoning-arena)
- **Process Reward Model**: [At1uri/financial-reasoning-process-reward-model](https://huggingface.co/At1uri/financial-reasoning-process-reward-model)
- **Research Vault**: [At1uri/quantitative-llm-reasoning-vault](https://huggingface.co/buckets/At1uri/quantitative-llm-reasoning-vault)
- **Flagship Collection**: [Institutional Quantitative AI Suite](https://huggingface.co/collections/At1uri/institutional-quantitative-ai-suite-6aeeaed0365774ae33b9621e)
---
## 📜 Citation
```bibtex
@dataset{atluri2026quantcot,
author = {Atluri, Ramprasad},
title = {Quantitative Reasoning Chain-of-Thought (CoT) Benchmarks},
year = {2026},
publisher = {Hugging Face},
howpublished = {\url{https://huggingface.co/datasets/At1uri/quantitative-reasoning-cot-benchmarks}}
}
```