Datasets:
Publish POSA Bench preview-v0.1.0
Browse files
README.md
CHANGED
|
@@ -21,6 +21,12 @@ size_categories:
|
|
| 21 |
> repetition and a single AI semantic reviewer. They are not official
|
| 22 |
> leaderboard ranks.
|
| 23 |
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 24 |
POSA Bench evaluates small, offline language models for the Portable Offline
|
| 25 |
Survival Assistant. Its 25 cases test source-grounded emergency guidance,
|
| 26 |
critical-fact recall, unsafe contradictions, citations, and appropriate
|
|
|
|
| 21 |
> repetition and a single AI semantic reviewer. They are not official
|
| 22 |
> leaderboard ranks.
|
| 23 |
|
| 24 |
+
[Open the live fixed-context preview](https://huggingface.co/spaces/UnpluggedAI/posa-on-device-leaderboard)
|
| 25 |
+
|
| 26 |
+
Current directional result: Liquid LFM2.5 has the strongest reviewed content;
|
| 27 |
+
Qwen3 has the fastest first-token latency. See `PREVIEW.md` for the complete
|
| 28 |
+
numbers, safety findings, controlled configuration, and claim boundaries.
|
| 29 |
+
|
| 30 |
POSA Bench evaluates small, offline language models for the Portable Offline
|
| 31 |
Survival Assistant. Its 25 cases test source-grounded emergency guidance,
|
| 32 |
critical-fact recall, unsafe contradictions, citations, and appropriate
|