offlineisbetter

inference for models by offlineisbetter.

we believe that you shouldn't give your data to faceless companies, that you deserve to run text models locally, and that you shouldn't need to buy expensive hardware. so we're building offlineisbetter.

offlineisbetter models will be for encoding tasks: sentiment analysis, text tagging, document retrieval, etc., rather than for decoding tasks like autoregressive generation. we believe that it's wasteful and dangerous to depend on cloud apis for frontier language models to do these simple tasks, and it should be almost mindless to download a model to use it without dealing with runtimes or quantization formats.

try it yourself

try our first model yourself. our first model is a small (230m) text model for sentiment analysis called offline-sentiment-small.

pip install offlinedemo

download the model archive from the releases page of this repo and unpack the model.

tar -xvf offline-sentiment-small.tar

run the demo, passing the inflated directory containing the model checkpoint.

offlinedemo offline-sentiment-small

benchmarks

below are benchmarks for offline-sentiment-small on binary sentiment classification using the stanfordnlp/sst2 validation set. all benchmarks were completed on the ryzen 9950x3d cpu on one thread.

philosophy

succinctly, the core philosophy of offlineisbetter is that parameter-efficient and low-latency models should be easily accessible to everybody. of course hugging face and transformers.pipeline allows you to run sentiment analysis in three lines of python, but for more parameter-efficient models, already quantized and with optimized computation graphs.