StephanAkkerman/wallstreetbets-ner

WallStreetBets NER Dataset

A high-quality, human-annotated Named Entity Recognition (NER) dataset composed of 671 gold-standard samples extracted from Reddit's r/wallstreetbets. This dataset is explicitly curated to train zero-shot and token-classification models (like GLiNER) to parse chaotic retail investor text, financial slang, and complex asset mentions.

---

📊 Dataset Overview

• Total Samples: 671 curated posts and long-form "Due Diligence" (DD) manifestos.
• Total entities: 4017 (6.0 per sample avg)
• Tickers: 1832
• Companies: 2185
• Avg text length: 1344 chars
• Target Domain: Retail finance, options trading, and market commentary.
• Format: Clean text paired with explicit absolute character offset annotations (start, end, label).

Entity Schema

Label	Description	Examples
ticker	A market ticker symbol representing a tradable financial asset/option underlying.	$AAPL, TSLA, AMD, GLD
company	The corporate body, business entity, brand, employer, or manufacturer.	Microsoft, Melvin Capital, NVIDIA

---

🎯 Annotation Guidelines & Boundary Rules

To ensure high precision and eliminate labeling noise, annotations strictly adhere to the following linguistic and data engineering boundaries:

1. The Context Disambiguation Rule (e.g., AMD / Tesla)

Entities are classified based on how they function in the sentence context, not solely by the keyword itself:

• Labeled as ticker: "Long on AMD 180c options" or "Buying shares of TSLA" (Functions as a transactional asset/market symbol).
• Labeled as company: "AMD released their new AI chips" or "I want to work at Tesla" (Functions as a corporate entity, manufacturer, or employer).

2. Strict Currency Exclusions (Noise Mitigation)

• Fiat Currencies: Monetary values, bills, and account units (e.g., $2, $22, $10k, USD, EUR) are hard-excluded and never annotated.
• Cryptocurrencies: Digital protocol tokens (e.g., BTC, ETH, SOL) are excluded from this equity-centric release to prevent semantic drift.

3. Commodities Disambiguation

• Generic Commodities: Words like "gold", "silver", or "oil" are treated as common nouns and left unlabelled.
• Commodity Instruments: Asset-tracking ETFs (e.g., GLD, SLV) are strictly labeled as ticker.

---

🛠️ Quick Start

You can load this dataset instantly using the Hugging Face datasets library:

from datasets import load_dataset

# Load the dataset
dataset = load_dataset("StephanAkkerman/wallstreetbets-ner")

# Preview a sample record
sample = dataset["train"][0]
print(f"Text:\n{sample['text']}\n")
print("Entities Found:")
for start, end, label in zip(sample["entities"]["start"], sample["entities"]["end"], sample["entities"]["label"]):
    entity_text = sample["text"][start:end]
    print(f" - [{label}] {entity_text} ({start} -> {end})")

🧬 Intended Use & Model Training

This dataset is designed for training small-footprint sequence models and fine-tuning adapters (like LoRA models for GLiNER2).

• Training Note: For long posts over 256 tokens, it is highly recommended to perform overlapping sliding-window chunking (e.g., 150-word chunks with a 40-word overlap) to fit attention windows and optimize VRAM allocation without chopping tokens on boundaries.

Repositories

In the repositories below you can find how I used this data myself.

Adapter Training: stock-recognizer-model Engine / Inference: stock-recognizer