README.md

---
language: en
license: apache-2.0
tags:
  - token-classification
  - ner
  - finance
  - energy
  - geopolitics
  - distilbert
  - multitask
pipeline_tag: token-classification
---

# DistilBERT Energy Intelligence Multitask NER — v2

**Model ID:** `Quantbridge/distilbert-energy-intelligence-multitask-v2`

A domain-specific fine-tuned [DistilBERT](https://huggingface.co/distilbert-base-uncased) model for Named Entity Recognition across **energy markets, financial instruments, geopolitics, corporate events, and technology**. This is a broad-coverage multitask NER model designed for intelligence extraction from financial news and market commentary.

The model recognises **59 entity types** (119 BIO labels including B-/I- prefixes) spanning multiple intelligence domains.

---

## Entity Taxonomy

### Financial Instruments & Markets
| Label | Description |
|---|---|
| `EQUITY` | Stocks and equity instruments |
| `DERIVATIVE` | Futures, options, swaps |
| `CURRENCY` | FX pairs and currencies |
| `FIXED_INCOME` | Bonds, treasuries, notes |
| `ASSET_CLASS` | Broad asset class references |
| `INDEX` | Market indices (S&P 500, FTSE, etc.) |
| `COMMODITY` | Physical commodities (oil, gas, metals) |
| `TRADING_HUB` | Price benchmarks and trading hubs |

### Financial Institutions
| Label | Description |
|---|---|
| `FINANCIAL_INSTITUTION` | Banks, brokerages, investment firms |
| `CENTRAL_BANK` | Central banks (Fed, ECB, BoE) |
| `HEDGE_FUND` | Hedge funds and asset managers |
| `RATING_AGENCY` | Credit rating agencies |
| `EXCHANGE` | Stock and commodity exchanges |

### Macro & Policy
| Label | Description |
|---|---|
| `MACRO_INDICATOR` | GDP, inflation, unemployment figures |
| `MONETARY_POLICY` | Interest rate decisions, QE programmes |
| `FISCAL_POLICY` | Government spending, tax policy |
| `TRADE_POLICY` | Tariffs, trade agreements, WTO actions |
| `ECONOMIC_BLOC` | G7, G20, EU, ASEAN, etc. |

### Energy Domain
| Label | Description |
|---|---|
| `ENERGY_COMPANY` | Oil majors, utilities, renewable firms |
| `ENERGY_SOURCE` | Oil, gas, coal, solar, nuclear, etc. |
| `PIPELINE` | Energy pipelines and transmission lines |
| `REFINERY` | Oil refineries and processing plants |
| `ENERGY_POLICY` | OPEC decisions, energy legislation |
| `ENERGY_TRANSITION` | Decarbonisation, net-zero, EV, hydrogen |
| `GRID` | Power grids and electricity networks |

### Geopolitical
| Label | Description |
|---|---|
| `GEOPOLITICAL_EVENT` | Summits, elections, geopolitical shifts |
| `SANCTION` | Economic sanctions and embargoes |
| `TREATY` | International agreements and accords |
| `CONFLICT_ZONE` | Active or historic conflict regions |
| `DIPLOMATIC_ACTION` | Diplomatic moves, expulsions, negotiations |
| `COUNTRY` | Nation states |
| `REGION` | Geographic regions (Middle East, EU, etc.) |
| `CITY` | Cities and urban locations |

### Corporate Events
| Label | Description |
|---|---|
| `COMPANY` | General companies |
| `M_AND_A` | Mergers and acquisitions |
| `IPO` | Initial public offerings |
| `EARNINGS_EVENT` | Quarterly earnings, revenue reports |
| `EXECUTIVE` | Named C-suite executives |
| `CORPORATE_ACTION` | Dividends, buybacks, restructuring |

### Infrastructure & Supply Chain
| Label | Description |
|---|---|
| `INFRA` | Physical infrastructure (general) |
| `SUPPLY_CHAIN` | Supply chain disruptions and logistics |
| `SHIPPING_VESSEL` | Named ships and tankers |
| `PORT` | Ports and maritime hubs |

### Risk & Events
| Label | Description |
|---|---|
| `EVENT` | General newsworthy events |
| `RISK_FACTOR` | Risk factors and vulnerabilities |
| `NATURAL_DISASTER` | Hurricanes, earthquakes, floods |
| `CYBER_EVENT` | Cyber attacks and digital incidents |
| `DISRUPTION` | Supply or market disruptions |

### Technology
| Label | Description |
|---|---|
| `TECH_COMPANY` | Technology companies |
| `AI_MODEL` | AI systems and models |
| `SEMICONDUCTOR` | Chips and semiconductor companies |
| `TECH_REGULATION` | Technology regulation and policy |

### People & Organizations
| Label | Description |
|---|---|
| `PERSON` | Named individuals |
| `THINK_TANK` | Policy research organizations |
| `NEWS_SOURCE` | Media and news outlets |
| `REGULATORY_BODY` | Government regulators (SEC, FCA, etc.) |
| `ORG` | General organizations |

---

## Usage

```python
from transformers import pipeline

ner = pipeline(
    "token-classification",
    model="Quantbridge/distilbert-energy-intelligence-multitask-v2",
    aggregation_strategy="simple",
)

text = (
    "The Federal Reserve held interest rates steady as Brent crude fell below $75 "
    "following OPEC+ production cuts and renewed sanctions on Russian energy exports."
)

results = ner(text)
for entity in results:
    print(f"{entity['word']:<35} {entity['entity_group']:<25} {entity['score']:.3f}")
```

**Example output:**
```
Federal Reserve                     CENTRAL_BANK              0.961
Brent                               TRADING_HUB               0.954
OPEC+                               REGULATORY_BODY           0.947
Russian energy exports              SANCTION                  0.932
```

### Load model directly

```python
from transformers import AutoTokenizer, AutoModelForTokenClassification
import torch

model_name = "Quantbridge/distilbert-energy-intelligence-multitask-v2"

tokenizer = AutoTokenizer.from_pretrained(model_name)
model = AutoModelForTokenClassification.from_pretrained(model_name)
model.eval()

text = "Goldman Sachs cut its oil price forecast after OPEC+ agreed to extend output cuts."
inputs = tokenizer(text, return_tensors="pt")

with torch.no_grad():
    outputs = model(**inputs)

predicted_ids = outputs.logits.argmax(dim=-1)[0]
tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])

for token, label_id in zip(tokens, predicted_ids):
    label = model.config.id2label[label_id.item()]
    if label != "O" and not token.startswith("["):
        print(f"{token.lstrip('##'):<25} {label}")
```

---

## Model Details

| Property | Value |
|---|---|
| Base architecture | `distilbert-base-uncased` |
| Architecture type | DistilBertForTokenClassification |
| Entity types | 59 types (119 BIO labels) |
| Hidden dimension | 768 |
| Attention heads | 12 |
| Layers | 6 |
| Vocabulary size | 30,522 |
| Max sequence length | 512 tokens |

---

## Intended Use

This model is designed for **financial and energy intelligence extraction** — automated NER over news feeds, earnings transcripts, regulatory filings, and geopolitical reports. It is a base model suitable for:

- Structured data extraction from unstructured financial news
- Entity linking and knowledge graph population
- Signal detection for trading and risk systems
- Geopolitical risk monitoring

### Out-of-scope use

- General-purpose NER on non-financial text
- Languages other than English
- Documents with heavy technical jargon outside the financial/energy domain

---

## Limitations

- English-only
- Optimised for news-style formal writing; may underperform on social media or informal text
- 59-label taxonomy may produce overlapping predictions for ambiguous entities (e.g. a company that is also an energy company)
- BIO scheme does not support nested entities

---

## License

Apache 2.0 — see [LICENSE](https://www.apache.org/licenses/LICENSE-2.0).