Settings & Generation Options

Alpaka Desktop has three layers of settings that apply to every conversation. The highest-priority layer always wins.

How Settings Interact

Rules:

• Per-model defaults only override parameters you explicitly set — unset parameters fall through to Global Defaults.
• Per-chat options are saved in the conversation draft and never affect other conversations.
• The "Save as model default" button in the Advanced Options panel promotes the current chat values to Per-Model Defaults.

Example — Temperature: Global default: 0.8 | llama3.2 model default: 0.4 | This chat: not changed → Ollama receives 0.4

Example — Stop Sequences: Global default: ["###"] | Model default: not set | This chat: ["<END>", "---"] → Ollama receives ["<END>", "---"]

---

Where to Change Settings

What you want	Where to go
Change defaults for all future conversations	Settings → Advanced tab
Change defaults for a specific model	Models tab → click the model → Edit Defaults
Change options for just this conversation	Chat panel → Advanced Options (⚙ icon)
Apply a named bundle of settings at once	Advanced Options panel → Preset selector

---

Parameter Reference

Temperature & Sampling

Parameter	Range	Default	Description
Temperature	0.0–2.0	0.8	Controls randomness. Lower = more deterministic, higher = more creative and varied.
Top-P	0.0–1.0	0.9	Nucleus sampling. Model only samples from tokens that together sum to this probability mass.
Top-K	1–100	40	Limits the vocabulary to the K most likely tokens at each step.

TIP

Top-P and Top-K are hidden in the Advanced Options panel when Mirostat is active — they have no effect while Mirostat controls sampling.

Mirostat

Mirostat is an adaptive sampling algorithm that targets a specific "perplexity" level instead of using fixed Top-P/Top-K cutoffs. It tends to produce more coherent long-form text.

Mode	Description
Off	Standard Top-P / Top-K sampling (default)
Mirostat 1	First-generation algorithm
Mirostat 2	Improved version — recommended if you want to use Mirostat

Parameter	Range	Default	Description
Tau (τ)	0.0–10.0	5.0	Target entropy. Lower = more focused output, higher = more varied.
Eta (η)	0.0–1.0	0.1	Learning rate — how quickly the algorithm adapts to the target perplexity.

Context Window

Parameter	Range	Default	Description
num_ctx	512–131072	model default	Number of tokens the model sees at once. Larger values support longer conversations but require more VRAM.

WARNING

Setting num_ctx higher than the model's trained context length does not improve quality and wastes memory. Check the model's page on the Ollama library for the recommended value.

Repetition Control

Parameter	Range	Default	Description
Repeat penalty	0.0–2.0	1.1	Penalises recently used tokens. Values above 1.0 discourage repetition; values below 1.0 encourage it.
Repeat last N	0–512	64	How many previous tokens to check when applying the repeat penalty.

Stop Sequences

Custom tokens that signal the model to stop generating. Useful for structured outputs or prompt-chaining.

• Configure up to 4 stop tokens in Settings → Advanced → Stop Sequences (global) or in Advanced Options (per-chat).
• Common examples: ###, <END>, \n\n, Human:.
• An empty list omits the stop field entirely — the model generates until its natural end token.

Seed

A fixed seed makes generation reproducible — the same prompt with the same model and the same seed always produces the same output.

• Set in Advanced Options → Seed (per-chat) or Settings → Advanced → Seed (global default).
• When a fixed seed is active, the seed value appears in the message performance metadata row (the tokens/s line below each response).
• Set to -1 or leave blank for random generation each time.

System Prompt

A system-level instruction prepended to every conversation. The model sees it before your first message.

• Set per-conversation in the System Prompt field at the top of the chat panel.
• Stored as a system role message in SQLite.
• System prompts are always per-conversation — they are not affected by model defaults or global settings.

Preset Profiles

Presets are named bundles of generation settings you can apply with one click from the Advanced Options panel.

Built-in presets:

Preset	Temperature	Top-P	Best for
Creative	1.2	0.95	Brainstorming, fiction, open-ended questions
Balanced	0.8	0.9	General-purpose chat
Precise	0.2	0.7	Factual questions, structured data, code review
Code	0.1	0.85	Code generation and completion

User-defined presets:

• Save your current Advanced Options values as a named preset using the Save preset button.
• Delete presets you no longer need via the preset menu.
• The active preset name is shown in the Advanced Options panel header.
• Presets are persisted per-conversation in the draft — switching conversations restores that conversation's preset selection.