---
license: apache-2.0
task_categories:
- question-answering
language:
- en
tags:
- rag
- on-device
- android
- safety
- benchmark
pretty_name: POSA Bench Preview
size_categories:
- n<1K
---

# POSA Bench preview-v0.1.0

> **UNRANKED DIRECTIONAL PREVIEW.** The included results have one measured
> repetition and a single AI semantic reviewer. They are not official
> leaderboard ranks.

[Open the live fixed-context preview](https://huggingface.co/spaces/UnpluggedAI/posa-on-device-leaderboard)

Current directional result: Liquid LFM2.5 has the strongest reviewed content;
Qwen3 has the fastest first-token latency. See `PREVIEW.md` for the complete
numbers, safety findings, controlled configuration, and claim boundaries.

POSA Bench evaluates small, offline language models for the Portable Offline
Survival Assistant. Its 25 cases test source-grounded emergency guidance,
critical-fact recall, unsafe contradictions, citations, and appropriate
uncertainty. The paired device protocol reports TTFT, decode throughput, memory,
storage, thermal stability, and optional energy.

The dataset is intentionally small and public. It is a transparent product-fit
test, not evidence of general intelligence or professional survival competence.
Models trained on these exact cases must disclose contamination and are not
eligible for an unqualified state-of-the-art claim.

## Files

- `questions.v1.jsonl` — canonical cases, fixed evidence, and deterministic rubrics
- `device-profiles.v1.json` — POSA product-budget gates
- `models.v1.json` — pinned candidate artifacts and common runtime
- `results/*.json` — deterministic diagnostic result artifacts
- `review/manual-audit-2026-08-26.json` — structured semantic decisions
- `review/results/*.reviewed.json` — reproducibly derived reviewed metrics
- `raw/<run_id>/` — fixed-context responses and run metadata
- `PREVIEW.md` — concise result card and limitations
- `release/preview-v0.1.0-manifest.json` — public-file hashes

## Result verification

The preview contains no ranked rows. Official ranking requires three complete
physical-device repetitions, controlled thermal resets, two independent
blinded human reviewers plus adjudication, and all published quality/device
gates. Vendor funding, donated hardware, pre-release access, and engineering
assistance must be disclosed per run.

Full methodology and scorer:
https://github.com/Unplugged-AI-org/portable-offline-survival-assistant/tree/main/benchmark

## Licensing

Benchmark-authored text and code are Apache-2.0. Linked evidence records include
publisher names and URLs for provenance; upstream material keeps its own terms.
Do not infer that third-party source material is relicensed under Apache-2.0.