account-123's picture shouc's picture
Initial commit
cecb5ee
|
Raw History Blame Contribute Delete
1.63 kB
metadata
license: mit
base_model: cantina-security/apex-flash-1
tags:
  - glm5_next
  - abliterated
  - security-research
pipeline_tag: image-text-to-text
library_name: transformers

apex-flash-1-abliterated

apex-flash-1-abliterated is an experimental derivative of apex-flash-1, Cantina Security's open-weights security model developed in partnership with Yeta (@yetalabs on X).

It is intended for authorized security research in environments the researcher owns or has permission to test. The variant modifies refusal behavior broadly; the changes are not limited to security tasks.

Evaluation

This variant has not undergone a separate full-suite evaluation. The results reported for apex-flash-1 apply to the standard checkpoint and should not be attributed to this derivative. Changes in refusal behavior do not establish improvements in security-task performance or reliability.

Capabilities

The checkpoint retains the standard model's image-text-to-text architecture. Image and video performance have not been evaluated for this variant.

Model details

The checkpoint uses the GLM-5.3-Flash architecture and is distributed in BF16. See the standard model card for the training overview and intended worker-model use.

Explore Apex to learn more about Cantina's security research system.

Read the Apex Flash release post for the training methodology and case examples.