A model 60 times smaller beat Google's FunctionGemma on Mobile Actions.
ARN: 4.4M parameters vs 268M. Same training data. 11 ms on one CPU thread.
The caveats, up front:
I looked at the test set while developing, so these results are exploratory.
With no rules, on a separate test, ARN reaches 40.9 on MixATIS and 70.3 on MixSNIPS. That is below several published models.
The full limitations are in the repo.
1
2
7
Model: huggingface.co/AnishPathak22…
Code: github.com/Anishp-cell/ARN
Paper: coming soon.
Feedback welcome, especially where the date rules or the fixed function list break on your data.
2
7
I just published Building a Persistent Codebase Memory System: Architecture, Hybrid Graph RAG, and Lessons Learned pub.towardsai.net/building-a…
21
I just published How I Built AETHER: A Local AI Assistant That Controls My PC, Sends WhatsApp Messages, and Learns… pub.towardsai.net/how-i-buil…
Thanks to @towards_AI for reviewing and accepting my article.
1
24