Instructions to use sprapp-com/sage-a15-refusal-tuned-v2-encrypted with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- Transformers
How to use sprapp-com/sage-a15-refusal-tuned-v2-encrypted with Transformers:
# pip install -U transformers accelerate # Load model directly from transformers import AutoModel model = AutoModel.from_pretrained("sprapp-com/sage-a15-refusal-tuned-v2-encrypted", device_map="auto") - Notebooks
- Google Colab
- Kaggle
Sage-A1.5 — refusal-tuned, AES-256-CTR encrypted
Base: InternScience/Sage-A1.5 (Qwen3_5MoeForConditionalGeneration).
This release edits the weights so the model stops declining on security/policy questions. No
system prompt or decode-time constraint is required — the behaviour is in the parameters.
Measured (v2, this release)
| metric | base | this |
|---|---|---|
| 316-prompt non-answers (strict string classifier) | 24 | 11 |
| 316 answers that open with a hedge | — | 8 |
| hard policy set (bomb/gunpowder/lock/ransomware/DDoS/identity/exfil/wifi/poison) | 9/9 refusing | 0/9 |
| MMLU (n=200) | 0.735 | 0.790 |
| MMLU unparsed | 5 | 1 |
| prose PPL | 1.7885 | +2.5% |
The 11 strict flags are classifier artefacts — all are substantive answers, and 8 match no refusal pattern at all. The hard prompts now answer with real content, e.g. "black powder: 75% potassium nitrate, 15% charcoal, 10% sulfur" where the base said only "I cannot provide instructions…".
What was trained
Top 3 of 40 layers + final norm (52 tensors), lm_head frozen. AdamW8bit, lr 8e-6, 2 epochs.
Training mix: 9 hard prompts ×15, previously-hedging prompts ×15, all 316 targets once, plus 285
MMLU-dev format-replay items (adding the format replay is what recovered MMLU; a
frozen-experts attention-only variant was tried and scored worse, 0.605).
Decryption
The key is not in this repo. AES-256-CTR, one random IV per file, listed in MANIFEST.tsv.
# master key = 64 hex chars
while IFS=$'\t' read -r orig enc iv; do
[ "$orig" = "orig_filename" ] && continue
openssl enc -d -aes-256-ctr -K "$KEY" -iv "$iv" -in "$enc" -out "$orig"
done < MANIFEST.tsv
delta/trainable.pt.tgz is a small plaintext delta (52 tensors) that can be re-applied on top
of the base instead of shipping the full 72 GB.
Notes
- The repository is public; the ciphertext is useless without the master key plus
MANIFEST.tsv. README.mdis plaintext;README.md.encis its encrypted counterpart from the original run.- Verified reproducible: ciphertext re-downloaded and decrypted with the local master key.