Keywords: character-level language model, curve encoder, dual-channel memory, word fingerprint, episodic slot memory, zero-train retrieval, canonical memory, knowledge editing, machine unlearning, RAG alternative, PyTorch, reproducible research.
TapeLM weights (variant A)
Checkpoints for TapeLM — one system: frozen dual-channel curve encoder (P1), matched GPT control, and optional family W for canonical fingerprint memory.
Weights: huggingface.co/Kostya03v/TapeLM-P1
Code: github.com/KonstantinK-V/TapeLM
Quickstart: artifact/QUICKSTART.md
Files
| File | Role |
|---|---|
stage191_p1_curve.pt |
P1 encoder — generation + fp/memory |
stage191_p2_gpt.pt |
Matched GPT control |
w_registry/* |
Optional family W (qmap read) |
Tokenizer: results/stage177_curve_bpe_tokenizer.json on GitHub.
Usage
git clone https://github.com/KonstantinK-V/TapeLM.git
cd TapeLM
pip install -r artifact/requirements.txt
python artifact/scripts/download_checkpoints.py
python artifact/scripts/run_product.py
License
MIT — TapeLM GitHub.