urbanflux-borough-nemotron3-nano

urbanflux-borough-nemotron3-nano is a fine-tuned London borough assignment model. It takes unstructured or semi-structured text as input and returns the most likely London borough name.

The model is designed for UrbanFlux-style data processing workflows where the borough may be implied by fields such as ward name, dataset metadata, CSV rows, source descriptions, population records, or other London civic data fragments.

What It Does

Given a text snippet, JSON object, CSV-like row, metadata block, or mixed urban data record, the model predicts the London borough represented by that input.

Example output:

Barking and Dagenham

Example

Input:

{
  "Ward_Name": "Whalebone",
  "Population": "19361",
  "Year": "2050",
  "source_metadata": "Land Area and Population Density, Ward and Borough",
  "task": "Identify the London borough represented by this CSV row."
}

Output:

Barking and Dagenham

Training Data

The model was fine-tuned using borough-assignment training examples from:

https://pub-f20eb55e72ee41a5b80036ea8f6107bb.r2.dev/urbanflux_borough_assignment_training.jsonl

The dataset contains instruction-style examples where the input includes London civic data records and the target output is the correct borough name.

Intended Use

This model is useful for:

• Assigning London borough names to messy or incomplete urban data
• Processing civic datasets where borough information is embedded in metadata
• Normalising borough labels in UrbanFlux pipelines
• Lightweight geographic classification for London-focused records

Output Format

The model is intended to return only the borough name, for example:

Camden

or:

Tower Hamlets

Model Details

• Architecture: nemotron_h_moe
• Parameters: 31.6B
• Quantization: Q4_K_M
• Context length: 4096
• Size: approximately 25GB

Run

ollama run chigwel/urbanflux-borough-nemotron3-nano:q4_k_m

Notes

The model predicts the most likely borough based on the information provided. If the input is ambiguous, incomplete, or refers to multiple boroughs, the output should be treated as a best-effort classification rather than a guaranteed geographic match.

Source used: training dataset JSONL.