Infactory Pulse 1 (4B)
Pulse 1 is a narrative intelligence model for extracting structured
storylines from news and editorial content.
What It Does
Given an article, Pulse 1 extracts storylines — not just topics or keywords,
but structured narratives that answer:
- WHO is acting (actor)
- WHAT they did (action)
- TO WHOM (target)
- WITH WHAT CONSEQUENCE (outcome)
Each storyline includes a concise name, a why explanation, and full
Semantic Role Labeling (SRL) decomposition in machine-readable JSON.
Why Storylines?
A storyline is more than a topic. Topics are static categories ("economy",
"politics"). Storylines are dynamic narratives that capture cause-and-effect,
developing situations, and active debates.
Example:
- Topic: "Federal Reserve"
- Storyline: "Federal Reserve interest rate increases slow U.S. housing market activity"
The storyline captures the actor (Fed), action (rate increases), target
(housing market), and outcome (slowdown) — information you can act on.
Model Details
| Property | Value |
|---|
| Base model | Gemma 3 4B (text-only) |
| Architecture | Gemma 3 (34 transformer layers) |
| Format | MLX safetensors |
| Precision | bfloat16 |
| Context window | 4,096 tokens |
| Size | ~8.5 GB |
Output Format
The model returns a JSON object with a topics array. Each storyline has six fields:
1{
2 "topics": [
3 {
4 "name": "Federal Reserve interest rate increases slow U.S. housing market activity",
5 "why": "The Fed's tightening cycle is making mortgages unaffordable for first-time buyers.",
6 "actor": "Federal Reserve",
7 "action": "increases interest rates",
8 "target": "U.S. housing market",
9 "outcome": "mortgage applications fall to decade lows"
10 }
11 ]
12}
| Field | Description |
|---|
name | A specific narrative phrase (6+ words) with concrete actor, action, and outcome |
why | One-sentence summary of why this storyline matters |
actor | The primary entity taking action (use full names, not abbreviations) |
action | What they did or are doing |
target | Who or what is affected |
outcome | The consequence or result |
Quick Start
With Ollama (recommended on Apple Silicon)
1# Import the model (from this directory after downloading weights)
2ollama create infactory_pulse1_4b -f Modelfile
3ollama run infactory_pulse1_4b
Or use the OpenAI-compatible API:
1curl -s http://localhost:11434/v1/chat/completions -d '{
2 "model": "infactory_pulse1_4b",
3 "messages": [
4 {"role": "user", "content": "Identify the key storylines discussed in this article.\n\nArticle:\nTitle: Fed Cuts Rates\nText: The Federal Reserve cut interest rates by 25 basis points today, citing slowing inflation and a cooling labor market."}
5 ]
6}'
Ollama runs Gemma 3 natively on Apple Silicon (MLX backend) and on CUDA for
NVIDIA hardware. The Modelfile sets num_ctx=4096, temperature=0.3,
top_p=0.9, and the appropriate Gemma 3 chat template.
With mlx_lm
For direct MLX inference:
1pip install mlx-lm
2python -m mlx_lm.generate \
3 --model ./ \
4 --prompt 'Identify the key storylines discussed in this article. ...'
Intended Use
Pulse 1 is designed for narrative intelligence workflows where you need to understand
not just what an article is about, but what is happening — the actors, actions, and consequences.
Content Intelligence Pipelines
- Storyline extraction — Convert unstructured articles into structured narrative data
- Salience scoring — Score sentences against extracted storylines to find the most relevant passages
- Entity resolution — Ground storylines in detected entities for richer metadata
- Semantic search — Index and retrieve content by narrative dimensions (actor, action, target, outcome)
Media Monitoring & Analytics
- Narrative tracking — Monitor how storylines evolve across publications over time
- Trend detection — Identify emerging storylines by aggregating across article streams
- Brand-adjacent content discovery — Find articles whose narratives align with brand themes
- Competitive intelligence — Track storylines mentioning specific companies, products, or people
Editorial & Publishing Workflows
- Automated tagging — Generate structured metadata for content management systems
- Pull quote extraction — Score article sentences to surface the most quotable passages
- Evergreen content discovery — Find archival articles newly relevant to today's storylines
- Newsletter curation — Cluster and summarize articles by shared narratives
Research & Analysis
- Narrative framing analysis — Study how different publications frame the same events
- Discourse mapping — Understand the actors and relationships in a topic area
- Information extraction — Build structured datasets from news corpora
Out-of-Scope Use
- General-purpose chat or instruction following
- Languages other than English
- Domains outside news and editorial content
- Real-time or safety-critical applications
Files
| File | Description |
|---|
model-00001-of-00002.safetensors | Model weights (shard 1/2) |
model-00002-of-00002.safetensors | Model weights (shard 2/2) |
model.safetensors.index.json | Weight shard index |
config.json | Model architecture configuration |
tokenizer.json | Fast tokenizer |
tokenizer_config.json | Tokenizer settings |
chat_template.jinja | Chat template (Gemma turn format) |
Modelfile | Ollama-compatible model definition |
License
This model is a fine-tuned derivative of Google's Gemma 3 and inherits the
Gemma Terms of Use.
- Permitted: Research, personal use, and commercial use in products and services.
- Required: You must agree to the Gemma Terms of Use before downloading or using
the weights. Redistributions must include the license terms.
- Prohibited: Use for unlawful purposes, generating harmful content, or
circumventing safety filters. The Gemma Prohibited Use Policy applies.
- Attribution: Derivative models must acknowledge Gemma as the base model.
The Infactory-specific fine-tuning is proprietary to Infactory AI.
Links
