coslinedev/financial-rlvr-1k-teaser

license: mit task_categories: - text-generation - reinforcement-learning language: - en tags: - finance - quant - rlvr - grpo - python - synthetic - data-wrangling size_categories: - 1K<n<10K

📈 Financial RLVR & Data Wrangling Benchmark Suite

1,000 Execution-Verified Trajectories for Zero-Hallucination Financial Reasoning

🔥 Empirical Proof • ⚡ Quickstart • 📊 Dataset Breakdown • 💼 Enterprise Edition

---

💡 Why this dataset exists: Standard open-source LLMs severely fail on financial code execution due to floating-point drift, unparsed messy feed strings ($, N/A), and mathematical hallucinations under zero-shot prompting. This dataset forces models to reason via executable Python code with strict AST & runtime verification.

---

🔥 Empirical Proof & Benchmark Results

We benchmarked leading open-weights base models before and after SFT / GRPO fine-tuning on this dataset suite.

Evaluation was measured using Pass@1 Accuracy (strict $1e-3$ floating-point tolerance + correct JSON payload execution) across 200 held-out test prompts.

📊 Pass@1 Performance Comparison

Model Architecture	Baseline Pass@1	Fine-Tuned (With Dataset) Pass@1	Trap Detection Rate	Pass@1 Delta
Qwen-2.5-Coder-3B-Instruct	34.2%	83.5%	94.0%	+144.1%
DeepSeek-R1-Distill-Qwen-8B	52.0%	89.2%	96.5%	+71.5%
Llama-3.1-8B-Instruct	38.5%	81.0%	91.2%	+110.3%

🎯 Key Empirical Takeaways

• 🛡️ 94%+ Anti-Overfitting Trap Success: Standard models hallucinate solutions even when given mathematically impossible inputs (e.g., negative time-to-maturity $T < 0$). Models trained on this dataset correctly trigger exception handling (TRAP_DETECTED).
• 🧹 Clean Feed Wrangling Rate: Jumped from 28.4% to 91.5% on raw, uncleaned currency inputs ("$1,250.50 USD", "N/A", "CORRUPTED").
• ⚡ Ultra-Compact Footprint: Fine-tuning on just 1,000 samples yields over 80%+ execution accuracy, proving high-density algorithmic dataset quality > raw token volume.

---

⚡ Quickstart

No authentication required. Load the dataset in 1 line of code:

from datasets import load_dataset

# Load 1,000 verified reasoning trajectories
dataset = load_dataset("coslinedev/financial-rlvr-1k-teaser")

# Inspect the first sample
print(dataset["train"][0])
Load via PandasPythonimport pandas as pd

url = "[https://huggingface.co/datasets/coslinedev/financial-rlvr-1k-teaser/raw/main/finquant_eval_1000_teaser.jsonl](https://huggingface.co/datasets/coslinedev/financial-rlvr-1k-teaser/raw/main/finquant_eval_1000_teaser.jsonl)"
df = pd.read_json(url, lines=True)
print(f"✅ Successfully loaded {len(df)} verified samples!")
📊 Dataset Architecture                                  [ 1,000 VERIFIED SAMPLES ]
                                              |
        +-------------------------------------+-------------------------------------+
        |                                     |                                     |
        v                                     v                                     v
📊 Quant Derivatives (70%)             🧹 Dirty Wrangling (15%)             🪤 Anti-Overfitting Traps (15%)
• Black-Scholes Pricing                • Raw String Data Cleaning           • Negative Maturity (T < 0)
• Greeks (Delta, Gamma, Vega)          • Missing Value Interpolation        • Negative Volatility (σ ≤ 0)
• Volatility Surface / Yields          • Unstructured Currency Feeds        • Corrupted Data Rejection
🔬 4-Stage Verification SandboxUnlike web-scraped text datasets, every single sample in this repository is 100% execution-verified. A sample receives total_reward = 1.0 ONLY if it passes all 4 sandbox checkpoints:[ Model Code Output ]
          │
          ├──► 1. AST Syntax Parser (20%) ────────► Ensures zero syntax/indentation errors
          ├──► 2. Quant Keyword Validator (25%) ──► Validates use of numpy, pandas, scipy, re
          ├──► 3. Python Execution Sandbox (30%) ─► Executes code in isolated runtime environment
          └──► 4. JSON Payload Matcher (25%) ────► Validates numerical precision within 1e-3 tolerance
          │
          ▼
   [ REWARD = 1.00 ]
🛠 Schema DefinitionFieldTypeDescriptionidstringUnique sample identifier (finquant-eval-0001)domainstringTask category (Options & Derivatives, Real-World Data Wrangling)is_edge_casebooleanTrue if the sample represents a logic trap/corrupted inputpromptstringThe instruction given to the modelcode_solutionstringExecutable Python script generating JSON outputexpected_payloaddictGround-truth JSON payload expected from executiontotal_rewardfloatVerification score (1.0 = 100% pass)JSON{
  "id": "finquant-eval-0042",
  "domain": "Real-World Data Wrangling",
  "is_edge_case": true,
  "prompt": "[Dirty Data] Calculate Volatility from invalid feed: ['N/A', 'CORRUPTED', '$--']",
  "code_solution": "import json\nfeed = ['N/A', 'CORRUPTED', '$--']\nvalid = [x for x in feed if x.replace('$','').replace('.','').isdigit()]\nif len(valid) < 2:\n    print(json.dumps({'status': 'TRAP_DETECTED', 'reason': 'Insufficient valid numerical data.'}))",
  "expected_payload": {
    "status": "TRAP_DETECTED"
  },
  "total_reward": 1.0
}
💼 Full Enterprise Dataset (50,000 Samples)This repository serves as the 1,000-sample Public Benchmark & Evaluation Teaser.For institutional AI labs, hedge funds, and LLM providers requiring our complete 50,000-sample Enterprise Dataset, custom execution sandboxes, proprietary data synthesis pipelines, or domain-specific dataset customization:📧 Contact Email: contact@cosline.dev🌐 Organization: Cosline Dev🤝 Use Cases: Institutional Quant Models, Automated Financial Analysts, Code-Execution LLM Tuning.