Pipeline Settings

This guide covers the PipelineSettings singleton — the app-level configuration for adaptation pipeline model defaults.

Overview

PipelineSettings is a singleton model (always pk=1) that provides default LLM model assignments for each pipeline node. These defaults apply to all adaptations unless overridden at the ad unit level.

Accessing PipelineSettings

Navigate to Core > Pipeline Settings in the admin
There is exactly one record — click to edit

If no record exists, one is created automatically when first accessed (via PipelineSettings.get_instance()).

Configurable Fields

Field	Pipeline Node	Description
Global Default Model	All nodes (fallback)	Used when no node-specific default is set
Concept Default Model	Concept Analyst	Extracts themes from origin script
Culture Default Model	Cultural Researcher	Produces cultural brief for target market
Format Gate Default Model	Format Evaluator	Checks language compliance
Culture Gate Default Model	Cultural Evaluator	Validates cultural sensitivity
Concept Gate Default Model	Concept Evaluator	Verifies concept preservation
Brand Gate Default Model	Brand Evaluator	Checks brand consistency

Note that the writer node is intentionally excluded — it defaults to the language’s primary model instead, since it produces target-language content. Override the writer via the ad unit’s pipeline_model_config if needed.

All fields are optional ForeignKeys to LLMModel. If left blank, the node falls through to the global default model, and then to the language’s primary model.

How Settings Interact with Overrides

PipelineSettings sits in the middle of the model resolution chain:

Ad Unit override        ← highest priority
  ↓
PipelineSettings node default
  ↓
PipelineSettings global default
  ↓
Language primary model   ← lowest priority

This means:

PipelineSettings sets the baseline for all adaptations
Ad unit overrides can change individual nodes for specific adaptations
Language primary model is the ultimate fallback if nothing else is configured

See Per-Node Model Selection for the full resolution chain and examples.

Configuration Strategies

Uniform Model

Set only the Global Default Model and leave all node-specific fields blank. Every node uses the same model:

Simple to manage
Good starting point
All nodes share the same quality/speed tradeoff

Split by Role

Assign different models based on node role:

Analysis nodes (concept, culture) — use a model strong in reasoning and cultural knowledge
Evaluation nodes (format, cultural, concept, brand gates) — use a model strong in structured evaluation
Writer node — handled by language primary model (no PipelineSettings field)

This balances quality where it matters most (evaluation gates) with efficiency elsewhere.

Quality vs. Speed

Approach	Model Choice	Tradeoff
Maximum quality	Qwen2.5-7B for all nodes	Slower, more VRAM, better evaluation accuracy
Balanced	7B for evaluators, 3B for analysis	Good quality with reasonable speed
Maximum speed	Qwen2.5-3B for all nodes	Fastest, lowest VRAM, may miss nuanced issues
4-bit quantized	Any model with `load_in_4bit=True`	Reduces VRAM by ~50%, slight quality loss

Editing in the Admin

Navigate to Core > Pipeline Settings
Click the settings record
For each node, select an LLM model from the dropdown (or leave blank for fallback)
Set the Global Default Model — this is the baseline for unconfigured nodes
Save

Changes take effect immediately for new adaptation tasks. Running tasks continue with their already-resolved models.

LLM Models

Available LLM models are managed under Core > LLM Models. Each model has:

Model ID — HuggingFace model identifier (e.g., Qwen/Qwen2.5-7B-Instruct)
Name — display name
Load in 4-bit — whether to use 4-bit quantization (requires bitsandbytes)
Is Active — only active models appear in PipelineSettings dropdowns

To add a new model, create an LLMModel record and it becomes available for selection in both PipelineSettings and per-ad-unit overrides.