"I've heard the community's feedback regarding the use of .pkl (Pickle) files in the AXIOM architecture. I'm happy to announce that we have officially migrated the Intent Classifier to .safetensors
Most of the thing was already safe it was just that the only project-specific file still using pickle is:
models/intent_model/setfit_intent_classifier/model_head.pkl (28KB).
- Why was that file still a .pkl?
This file contains the Classification Head (a Scikit-learn LogisticRegression model).
Reason: SetFit (the library being used for intent classification) uses Scikit-learn by default. Scikit-learn does not natively support .safetensors because it's designed for deep learning tensors, not traditional machine learning models.
The Default Behavior: When called model.save_pretrained(), SetFit pickles this small "head" model while saving the "body" as safetensors.
Key Improvements:
Enhanced Security: Eliminated the risk of arbitrary code execution inherent in Pickle files. The model head now uses a custom zero-copy inference engine with
safetensors
.
Pure Tensor Math: We now perform manual matrix-based predictions, which completely decouples the runtime from Scikit-Learn (reducing binary footprint).
Verified Parity: Extensive testing confirms 100% accuracy parity with the previous model—zero drift in confidence scores.
The changes are now live on the main branch. You can check out the full
Migration Plan in the repository for technical details.
Thanks to everyone who flagged this—keeping AXIOM secure and high-performing is our top priority!"
THANKS!! FOR 500+CLONES WITHIN 48 HRS..
TO MAKE IT VISIBLE TO LARGE COMMUNITY PLS GIVE A STAR TO THE REPO IT WILL KEEP US CONNECTED AS WELL
I dont need stars as for fame or smthing It just that github algorithm works like that .. I am happy with this high utility of the repo