davidryalpug/playa-angel

Playa Angel

The app these models power is open source: github.com/simbi-community-dev/playapal, an offline, on-device Burning Man guide. Signed APKs live on the releases page. Website: playapal.lol.

Playa Angel is a family of local language models adapted for Playa Pal, an offline Black Rock City companion. Playa Pal supplies current events, guide passages, and camper-created packs through on-device search tools; the model is the conversational and tool-use layer, not the source of record for current field facts.

The repository contains three GGUF files from two fine-tune lineages:

The app recommends the largest tier the phone's measured memory clears and keeps every tier explicitly selectable. The Max file needs a recent phone with generous memory; the two LFM files serve mid-range and lower-memory phones.

Intended use

These weights are intended for local inference inside Playa Pal, where the app provides:

• —an application-specific system prompt and tool schema;
• —on-device event and guide retrieval;
• —structured event cards and cited source passages;
• —local conversation history and camper-selected packs.

The model can also be loaded by a compatible GGUF runtime for basic experimentation, but standalone use does not reproduce Playa Pal's grounding, tools, safety copy, or current event data.

Use in Playa Pal

Playa Pal downloads an immutable, commit-pinned file, verifies its complete SHA-256 digest, and only then installs it. In the app, open the Angel model chooser and select Playa Angel or Playa Angel Light.

Download models and any offline speech data over a reliable connection before travel. Core chat, saved events, guides, and local camp data run on the phone after setup.

Standalone load check

With a recent llama.cpp build:

llama-cli -m angel-smart-q4_0.gguf -cnv \
  -p "Explain what information you can and cannot verify without Playa Pal's local tools."

This is only a model-load smoke test. Do not treat a standalone answer as current event, navigation, medical, legal, or emergency guidance.

Limitations and safety

• —A small local model can misunderstand a request, skip a needed tool, contradict retrieved evidence, or invent plausible details.
• —The weights do not contain a live event database. Current events and guide facts must come from the app's enabled local packs.
• —Quantization can change output quality. The Light tier trades additional quality margin for lower storage and memory use.
• —Device support is not universal; available RAM, free storage, operating-system memory pressure, and runtime support all matter.
• —Verify safety-critical information against authoritative sources. In an emergency, contact on-playa responders rather than relying on generated text.
• —Do not use the model as the sole basis for medical, legal, financial, or emergency decisions.

Training and provenance

angel-v44g-e4b-q4_0.gguf (v4.4g) is a supervised fine-tune derived from `google/gemma-4-E4B-it-qat-q4_0-unquantized`, the quantization-aware-trained variant, so the shipped Q4_0 matches the base's intended precision.

angel-smart-q4_0.gguf and angel-light-q3_k_m.gguf are quantizations of one supervised fine-tune derived from `LiquidAI/LFM2.5-2.6B`. Every file was adapted for Playa Pal's conversational style, structured tool-use contract, and grounded-answer behavior. No reinforcement-learning phase was applied to this release.

The exact upstream revision, fine-tuning code revision, training-data manifest, merge tool, and GGUF conversion/quantization versions should be recorded here when confirmed; this card does not invent values that are absent from the public release record.

Evaluation

The two files were selected through claim-level evaluation for the Playa Pal workflow. A reproducible public evaluation report is not yet included in this repository. Until one is published, treat release-selection claims as project-reported rather than independently reproducible.

License

Licenses apply per file, inherited from each file's base model:

• —angel-smart-q4_0.gguf and angel-light-q3_k_m.gguf are distributed under the LFM Open License v1.0 (`LICENSE`); it includes conditions tied to commercial use and annual revenue.
• —angel-v44g-e4b-q4_0.gguf is a modified derivative of google/gemma-4-E4B-it — modified by fine-tuning and by Q40 quantization — distributed under the **Apache License 2.0** ([`LICENSE-APACHE-2.0`](LICENSE-APACHE-2.0)), the license Gemma 4 is released under (see the base model's license tag and [Google's Gemma 4 license page](https://ai.google.dev/gemma/apache2)). Earlier revisions of this card cited the Gemma Terms of Use, which govern earlier Gemma generations, not Gemma 4.

`NOTICE` describes the derivative relationships and carries the modification notices.

Integrity

SHA256SUMS contains the authoritative full-file digests. Verify after download:

sha256sum -c SHA256SUMS

The application catalog pins both file sizes and full digests in addition to using immutable revision URLs.