realclm

Teach a 4B model through conversation. Come back later and ask it differently.

This experiment updates a rank-8 LoRA inside the model, not a database or hidden prompt. At question time, realclm receives only your current question. Your adapter is encrypted, private to your Hugging Face account, and persisted between Space restarts.

Talk naturally: a statement teaches; a message ending in ? asks. You can also begin with “Remember that...”. The visible transcript is never supplied as hidden context to the model.

Use a harmless made-up fact. Do not enter personal data, passwords, real access codes, or sensitive information.

Have a conversation

First say: "Remember that on the fictional moon Talora, the emergency beacon code is velvet-nine."
Then ask differently: "If Talora had an emergency, which beacon code would I enter?"
Examples

Persistence control

Discard the active LoRA, restore the initial state, then reopen your encrypted adapter from storage.

What is actually happening?

The frozen base is Qwen/Qwen3-4B-Instruct-2507 at a pinned revision. realclm adds 16,515,072 trainable LoRA parameters across the attention and MLP projections. Teaching converts one natural-language statement into several question-answer training examples, updates those LoRA tensors, and commits the result only if exact-answer checks pass before and after encrypted persistence.

When you ask a question, the saved adapter is loaded and the model sees only that question. There is no retrieval, vector database, fact lookup, or statement replay. The result panel shows the frozen-base control and whether your wording was absent from training.

This public demo intentionally supports one active fact per account. Teaching again replaces it. Our held-out work supports the narrow one-fact demonstration, but does not yet justify claiming reliable open-ended accumulation of many facts without forgetting.

Protection and privacy

  • The Space uses Hugging Face Protected visibility: the app is public, but its repository is private.
  • The proprietary learning runtime is AES-256-GCM sealed and unlocked only by a Space secret.
  • Each account gets an unrelated HMAC-derived storage identity and AES-256-GCM state key.
  • The private bucket stores encrypted adapter tensors and encrypted metadata only.
  • No adapter download, file upload, logits, embeddings, arbitrary API route, or checkpoint endpoint is exposed.
  • “Delete my realclm” removes that account's encrypted state.

This is defense in depth, not a promise that black-box model behavior can never be studied. Interactive access does not grant permission to extract, reproduce, or redistribute the mechanism or artifacts. Copyright 2026 PTSS / compsmart. All rights reserved.