Early 7B gate SFT and outcome RL evaluated checkpoints, with the exact frozen reader.
Archived on 2026-09-28 at the owner's request. Only evaluated checkpoints are included; calibration-only and rejected candidates are explicitly labelled.
| Location | Evaluation | Notes |
|---|---|---|
checkpoints/sft-epoch-1 | SFT epoch-end calibration | |
checkpoints/sft-epoch-2 | SFT epoch-end calibration | |
gate_adapter | outcome RL calibration | Default selected checkpoint; threshold 0.06 |
checkpoints/rl-step-64 | outcome RL calibration | |
checkpoints/rl-step-96 | outcome RL calibration | |
checkpoints/rl-step-128 | outcome RL calibration |
The root contains the exact frozen reader. Gate adapters are separate: do not merge a gate into the reader used for SKIM, memory updates and FINAL answers.
The default gate_adapter/ is early RL32 or position-balanced RL960, respectively. Other evaluated gates are under checkpoints/.
Download the entire repository with huggingface_hub.snapshot_download, then run python code/configure.py to set machine-local paths.
Use the included gate and recurrent reader code; a generic text-generation pipeline alone does not perform selective reading.
Historical gate metadata retains the original base identity and gate bias. Explicitly load the downloaded root reader and the selected adapter directory.
Early SFT128 has calibration BCE only and no selected inference threshold; thresholds must not be inferred from its filename.
Optimizer states are included where they were present locally. They are archival training files, not needed for inference.
See archive_manifest.json for checkpoint provenance, evaluation_records/ for available reports, and sha256_manifest.json for file integrity.
The archived inference wrapper accepts --adapter /absolute/path/to/checkpoints/NAME; early gates accept --threshold when explicitly needed. Adapter configuration base paths were relocated to this archive; adapter tensor bytes and gate bias are unchanged.