Ayushnangia/subliminal-learning-adapters

Subliminal Learning adapters

A public mega-repository of 491 unique PEFT adapter bundles from the recorded Subliminal Learning archive on Killarney. It contains adapters, configs, recorded tokenizer files and provenance documentation. It excludes base-model weights, intermediate checkpoints, smoke outputs, optimizer/RNG states and datasets.

Find an adapter

Browse adapters/channel/base-model/experiment/seed-or-scale/. ADAPTER_MAP.csv maps all 862 final-directory candidates to the 491 uploaded bundles, retaining every original source path and exact-copy alias. ADAPTER_INDEX.json records upload paths, byte sizes and SHA-256 hashes. Each adapter has its own README and original config under provenance/. A source-... suffix disambiguates different bundles with the same logical run name.

Loading

Use the matching base model with PEFT, selecting the desired adapter subfolder or downloading that folder first. The portable adapter_config.json identifies its recorded base model and immutable revision when known. Base weights must be obtained separately. Tokenizer files are included when present in the source. These are adapters, not standalone full models.

Evidence and limitations

Source directories classify number versus math channel; this is not an independent check of teacher/training provenance. Original configs are preserved. Portable configs change only the base-model path and a revision established by a recorded snapshot path. Unsloth namespaces are retained without claiming equivalence to upstream checkpoints. Only 16 of the 862 source candidates contain an immutable base revision in this evidence.

All 862 candidate safetensors files passed the file-size/data-offset check and were hashed; 371 were exact duplicate bundles. Of the 491 unique bundles, 207 have nearby trainer states reaching configured max_steps and 284 lack that completion evidence. Even a completed trainer state does not prove final bytes match its checkpoint. Unverified candidates are included as archived adapters with their uncertainty stated, not as scientifically validated runs.

This upload does not validate model quality, political scorer accuracy, statistical uncertainty, consumed training dose or safety. Research ownership differs from storage ownership. The original number-channel and math-channel review boundaries remain in force. Public visibility was authorized by the repository owner after the private storage allowance was exhausted. No license is asserted by this archive; users must check the applicable base-model and source permissions.

Upload complete: 491 adapter weights verified against source SHA-256 hashes; all 4,738 expected files present. Files were uploaded directly from Killarney.