TapeLM P1 + matched GPT (Stage 191)
Weights for the TapeLM variant A artifact: frozen dual-channel curve encoder (P1) and matched GPT-2 control used in Stages 192โ212.
Code & verdicts: github.com/KonstantinK-V/TapeLM
Architecture: docs/ARCHITECTURE.md
Files
| File | Role | ~Size |
|---|---|---|
stage191_p1_curve.pt |
P1 curve encoder (SelfModelXL d256, 6L, ~150M char pretrain) | 29 MB |
stage191_p2_gpt.pt |
Matched GPT control for fair comparisons | 22 MB |
stage177_curve_bpe.pt |
Optional BPE-on-curve checkpoint (177 path) | 4 MB |
BPE tokenizer JSON lives in the GitHub repo: results/stage177_curve_bpe_tokenizer.json.
Usage
git clone https://github.com/KonstantinK-V/TapeLM.git
cd TapeLM
pip install torch tokenizers transformers huggingface_hub
python artifact/scripts/download_checkpoints.py
python artifact/scripts/run_demo.py
Or download manually from this repo into checkpoints/:
checkpoints/stage191_p1_curve.pt
checkpoints/stage191_p2_gpt.pt
License
MIT โ same as TapeLM GitHub.
Citation
See CITATION.cff on GitHub.
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐ Ask for provider support