Hugging Face's logo

will-taggart
/
AFFECTOR-ARC

PyTorch
English
activation-steering
research
cuda
Model card Files
xet
 Community
AFFECTOR-ARC / macos /README.md
will-taggart's picture
will-taggart
Release macOS 1.0.4: exclude prompt and source wrappers from document output
953445e verified
preview code
|
Raw History Blame Contribute Delete
2.53 kB

AFFECTOR ARC for macOS 15+ — version 1.0.4

Native SwiftUI frontend for Apple Silicon Macs. It manages SSH, Hugging Face source updates, Slurm GPU setup, document rewriting, shape conversion/editing, job monitoring, and output downloads.

  • AFFECTOR-ARC-macOS15.zip: unpack and open AFFECTOR ARC.app. This local build is ad-hoc signed.
  • AFFECTOR-ARC-Desktop-source.zip: full editable Swift project, tests, remote agent, and build instructions.
  • remote_agent.py: stdlib SSH request agent bundled with the app; no listening server.

Changes in 1.0.4

Separate an explicit rewritten/attempt section from source/prompt wrappers emitted inside a model response. Exclude reasoning blocks, trailing diagnostics, and known prompt-commentary-only responses from the document field. Keep raw responses in the report. Verified against job 8680066 and inspected the installed app's displayed text.

Changes retained from 1.0.3

  • Keep TINKERCLIFFS CONNECTED fixed while connected, including during job polling.
  • Retain transformed text when changing tabs or job-list selection.
  • Restore the previous job/output on reconnect, and recover the newest rewrite when no previous job is saved.
  • Transfer generated text separately from the full calibration basis; keep full weights in downloadable reports.
  • Show generated passages only in Document and text exports; exclude retained originals, unprocessed source, prompts, targets, and report metadata.
  • Keep diagnostics and review counts in Jobs + results. The transformed-text field stays empty until generated text is available.
  • Use a normal-width system font for ARC at the same size and weight; preserve AFFECTOR.

The installed app was visually verified with a completed ARC rewrite. Swift core and result-recovery tests and all 14 remote-agent tests pass. Initial live GPU setup passed the seven upstream tests on an NVIDIA A100-SXM4-80GB, and a sample rewrite completed successfully.

Connection

Defaults: TinkerCliffs login node 1, SSH key ~/.ssh/arc_ed25519 with IdentitiesOnly=yes. The profile is prefilled for wtaggart and allocation will_taggart_mcll; change these for your own account. Password + Duo is also available. Credentials and private keys are not included in either archive.

After connecting, install/update the HF source if needed, prepare the GPU environment, import a document and emotional shape, then choose Transform on ARC. Existing GPU setup and completed jobs are reused. Downloads retain input, output, reports, geometry, and logs.