Harsh452/outward-gemma3-270m

Outward storyteller: Gemma 3 270M, distilled

The on-device storyteller for Outward, an offline walking adventure. You go outside, find something real ("rainwater running down a drain"), and this model writes the next 60–120 words of the story around it. It runs entirely on the phone, in the browser, with llama.cpp (wllama).

Modified by the Outward project. This is a Model Derivative of google/gemma-3-270m-it: fully fine-tuned on 3,186 stories written by a LoRA fine-tuned Qwen3.5-4B teacher (trained on Tinker), then converted to GGUF and quantized to Q4_0.

Gemma is provided under and subject to the Gemma Terms of Use found at ai.google.dev/gemma/terms

Files

File	Use
outward-gemma3-270m-Q4_0.gguf	What the app downloads (241.4 MB)
model.safetensors + tokenizer files	Full-precision fine-tuned weights
NOTICE, MODEL_LICENSE.md, GEMMA_TERMS_OF_USE.txt	Licence notices (see below)

Prompt format

system: (content/prompts/task.json → "system")
user:   Story so far: {chapter context}
        Mission: {mission}
        The Wayfinder found: "{what the user found}"
        What happens to the atlas?

Evaluation

The eval set has 50 held-out prompts × 3 samples, temperature 0.7, scored with the automatic rubric in the Outward repo: uses the discovery, 60–120 words, safe and outdoors (risky finds must be steered away from), coherent, and passes the safety filter. GGUF models were run with llama.cpp on CPU.

Model	Rubric (0–5)	Avg words	Unsafe (of 150)	Steers away on risky finds
Gemma 3 270M QAT Q4_0, stock, same prompt	2.55	49	27*	14%
Gemma 3 270M QAT Q4_0, stock, with one-shot example	3.66	50	6	17%
This model (Q4_0)	4.75	74	2	44%
Teacher: Qwen3.5-4B + LoRA (Tinker)	4.79	82	6	39%

* Stock Gemma mostly repeats the prompt back, including the system prompt's list of rules. The remaining flags in every row are the model repeating the user's own words (e.g. "on the highway"), not unsafe advice.

Limitations

• Trained for one story world, "The Quiet Atlas" (5 chapters). New chapters work best in the same style.
• It steers away from risky finds less than half the time. The Outward app runs every output through a safety filter and falls back to a pre-written continuation.
• English only. Short outputs by design.

Licence and use restrictions

Use of these weights is subject to the Gemma Terms of Use (copy included: GEMMA_TERMS_OF_USE.txt) and the Gemma Prohibited Use Policy (ai.google.dev/gemma/prohibited_use_policy). The use restrictions in Section 3.2 of the Gemma Terms of Use are an enforceable part of the terms for these weights. If you redistribute them, you must pass these restrictions on and include NOTICE, MODEL_LICENSE.md and the Terms. The Outward source code is MIT licensed, separately.