---

license: apache-2.0

base_model: Qwen/Qwen3-0.6B-Base

pipeline_tag: text-generation

language:

- en

tags:

- gguf

- qwen3

- sleep

- bedtime-stories

- storytelling

- text-generation

- browser

- wllama

- llama.cpp

---

dream-0.6b — a bedtime storyteller that fits in a browser tab

A 0.6B-parameter model that writes slow, gentle, plotless sleep stories for

grown-ups. Built to run fully client-side in the browser (via

wllama / WebAssembly) as the story engine

behind sleepcast.pro's dream feature — one ~400MB

download, then infinite offline stories with nothing leaving the device.

What it does

Give it a scene ("a greenhouse in soft rain", "a night ferry crossing calm

water") and it wanders: soft sensory prose, no plot, no urgency, no

cliffhangers — words designed to be listened to while falling asleep.

Prompt format

This is a completion model (not chat). Use exactly this frame, including

the primed opening line — the model continues from it:

A slow, gentle bedtime story for grown-ups. No plot, no urgency — soft words that wander until the listener drifts off.

Topic: {your scene}

Story:
Let yourself sink down into the quiet. Wherever you are right now, let it hold you. There's nowhere to be and nothing to do. Just breathe, slow and easy, and let these words carry you somewhere soft and far away.

Recommended sampling: temperature 0.8, top_p 0.95. The full intro prime

matters — shorter primes let the base model's multilingual pretraining leak

through at the seam.

llama.cpp

llama-completion -m dream-0.6b-Q4_K_M.gguf -p "$(cat prompt.txt)" \
  -n 2000 --temp 0.8 --top-p 0.95

LM Studio

Import the GGUF, use completion/playground mode with the frame above.

How it was made

Knowledge distillation from a two-stage local pipeline:

1. A Qwen3-4B-Base + QLoRA adapter (trained on sleep-podcast-style

narration) generated 120 raw stories across a 30-place × 8-texture topic

grid.

1. Gemma rewrote each one clean — consistent register, fixed

intro/outro brackets, no artifacts — yielding 119 records (~7.2k chars

each).

1. Qwen3-0.6B-Base was LoRA-fine-tuned (r32, α64, 8 epochs, bf16) on

the clean set, merged, and quantized to Q4_K_M with llama.cpp.

The entire pipeline ran on a single 8GB consumer GPU.

Files

v2 (current) — trained on a 300-story, 60-topic variety-first dataset;

real topic-following (v1 mode-collapsed toward one generic story). v2 has

NO fixed intro: the prompt ends bare at Story: and the first few tokens

may stutter — trim short leading lines. Use repeat penalty ~1.1.

| file | size | use |

|---|---|---|

| dream2-0.6b-Q4_K_M.gguf | ~397MB | browser (wllama), llama.cpp, LM Studio |

| dream2-1.7b-Q6_K.gguf | ~1.4GB | higher quality; Q4 breaks this one, use Q6 |

| dream2-4b-chat-Q8_0.gguf | ~4.3GB | desktop chat: send a topic as the user message |

| adapters/dream2-*.lora.gguf | 40-130MB | attach to stock Qwen3 bases via --lora |

4B chat system prompt: *"You are a bedtime storyteller for grown-ups. When

given a topic, tell a slow, gentle, plotless story in soft wandering words

until the listener drifts off. Never break character, never explain, just

tell the story."*

v1 (legacy)

| file | size | use |

|---|---|---|

| dream-0.6b-Q4_K_M.gguf | ~397MB | v1: single-voice, weak topic-following; needs the primed intro (see git history) |

Limitations

• English only; single register (soporific narration). It will not answer

questions, follow instructions, or hold a conversation — by design.

• No chat template training; chat-mode use produces off-register output.
• Small model: occasional soft non-sequiturs. In its intended use

(drowsy listening) these read as features.

• Not for factual, medical, or advisory content of any kind.

Lineage & license

Base model Qwen/Qwen3-0.6B-Base

(Apache 2.0). Fine-tune and distillation data produced by the sleepcast

pipeline.