Local Language Model

A locally-hosted model for narration context and the Seismic Chat interface — no third-party AI provider.

Runtime: Ollama

Intensity Lab uses a locally-hosted open-source language model for two distinct purposes: generating geographic context sentences for animation audio narration (see Text-to-Speech Narration), and powering the Seismic Chat interface where visitors can ask questions about seismology. All inference runs on the Intensity Lab server — no data is sent to any third-party AI provider.

Ollama is an open-source tool for running large language models locally. It manages model downloads, quantization, and a local HTTP inference API compatible with the OpenAI chat completions format. If Ollama is unreachable, audio narration falls back to generic text and the chat endpoint returns HTTP 503.

Model

Property	Value
Model	Meta Llama 3.2
Tag	llama3.2:latest
Parameters	3.2 billion
Quantization	Q4_K_M
Context length	128k tokens
System prompt	Seismology expert assistant; responses scoped to earthquake science

The version and quantization in use is recorded here deliberately: as newer local models become available, this page's baseline test (below) is the reference point for comparing response quality against a fixed, standardized question list before switching.

Uses

Animation audio narration — city enrichment. When generating a narration track for a simulated animation, the model is asked to supply the county name, state classification (city/town), and a notable characteristic for the affected city. This geographic context is validated and normalised before being incorporated into the two-sentence narration script read aloud by the TTS engine. If the model returns an unusable response the narration falls back to generic regional text.
Seismic Chat. The /chat/ page accepts free-text questions about seismology. Each question is submitted to the model with a system prompt that scopes responses to earthquake science. The chat interface is for general educational questions — all intensity estimates and earthquake data on this site are computed deterministically from USGS ShakeMap and the methods described elsewhere on this page, not generated by the language model.

Standardized Question List & Baseline Results

A fixed set of ten seismology questions is used to test any candidate model before it's deployed — the same questions are re-run against new model versions so results are comparable over time. The run below is the current baseline; the test suite checks only that each response is a non-empty string, so the responses themselves are for reading, not automated grading.

Field	Value
Test date	2026-06-28
Website version	d53a31e
Test suite	tests/test_chat.py (10 questions)
Result	10 / 10 passed
Total response time	12.39 s

Q1: What is the difference between seismic magnitude and seismic intensity?

Seismic magnitude refers to the size of an earthquake as measured by its amplitude or the distance it travels from the epicenter, usually expressed in logarithmic units such as Richter scale (MS) or moment magnitude scale (Mw). Seismic intensity, on the other hand, describes the actual effects experienced during an earthquake, including ground shaking, damage, and hazard. While magnitude measures the earthquake's size, intensity reflects its impact on the ground and surrounding structures.

Q2: What is the Richter scale?

The Richter scale was a logarithmic scale used to measure the magnitude of earthquakes. Developed by Charles F. Richter in 1935, it measured the amplitude of seismic waves recorded on seismographs at the moment of an earthquake. The scale ranged from 0 (microearthquakes) to around 10 (extremely powerful events). However, it has largely been superseded by more accurate and comprehensive magnitude scales like the Moment Magnitude Scale (MMS).

Q3: What is a Vs30 site condition?

Vs30 refers to the site factor of shear wave velocity at 30 Hz frequency, which represents the average shear wave speed over a depth of about 3 meters. It's used to account for the effects of local geology on seismic ground motion. A higher Vs30 value indicates a more stable and less deformable soil or rock condition, resulting in reduced shaking. Vs30 is an important parameter in seismic hazard assessment and building design.

Q4: What different seismic intensity scales exist and which is used in the United States?

There are several seismic intensity scales used worldwide, including the Modified Mercalli Intensity Scale (MMI), the Richter Magnitude Scale, and the Surface Wave Magnitude Scale. In the United States, the MMI scale is commonly used in ShakeMap products from the National Earthquake Information Center (NEIC) and other agencies.

Q5: What are some typical observed conditions for seismic intensity level VI?

A seismic intensity level 6 is considered strong to severe. Typically, people can feel the earthquake strongly indoors and outdoors, causing objects to fall off shelves and heavy furniture to vibrate or shift. Outside, trees may sway violently, power lines may break, and windows may shatter due to air pressure changes. The shaking also tends to be long-lasting, often several seconds to a minute.

Q6: How large can earthquakes get?

The largest earthquakes, known as Great Earthquakes or Megaearthquakes, have magnitudes between 8 and 9.5 on the moment magnitude scale (Mw). The largest recorded earthquake was the Great Chilean Earthquake of 1960, which had an estimated Mw of approximately 9.5.

Q7: What level of shaking is expected from a Magnitude 7 earthquake?

A Magnitude 7 earthquake typically produces moderate to strong shaking, often with peak ground acceleration (PGA) values ranging from 0.1 to 0.3 g, which can cause damage to buildings and other structures.

Q8: What is an earthquake focal mechanism?

An earthquake focal mechanism describes the orientation of the rupture plane and the direction of displacement at its center during an earthquake. It consists of three components: a strike-slip, dip-slip, or oblique faulting mechanism (the orientation of the plane); the normal or reverse faulting type; and the slip direction or amount on each component. This description helps predict an earthquake's potential impact on local terrain and infrastructure.

Q9: What is a ground motion prediction equation?

A Ground Motion Prediction Equation (GMPE) is a statistical model used to estimate the expected ground motion of an earthquake at a specific location. It is based on historical data, seismic hazard maps, and site characteristics. GMPEs are commonly used by seismologists to forecast shaking intensities and duration for a given earthquake scenario, and inform building codes, emergency preparedness planning, and infrastructure design.

Q10: Which ground motion prediction equations are commonly used in California and Alaska?

In California, the most commonly used GMPEs include equations developed from southern and central California data, specifically designed to predict strong ground motions along the San Andreas Fault. In Alaska, commonly used GMPEs were developed by the Pacific Earthquake Engineering Research Center (PEER) and the Federal Emergency Management Agency (FEMA), based on data from Alaska's active fault zones. Note: The specific model names in this response reflect Llama 3.2's training data, which may not perfectly match current USGS/NSHMP practice. For authoritative GMPE recommendations see the USGS National Seismic Hazard Model documentation.

As new local models become available, they will be tested against this same question list and the results compared here.