Instructions to use Observer04/abliteration-margin-audit-t22 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use Observer04/abliteration-margin-audit-t22 with PEFT:
from peft import PeftModel from transformers import AutoModelForCausalLM base_model = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global") model = PeftModel.from_pretrained(base_model, "Observer04/abliteration-margin-audit-t22") - Notebooks
- Google Colab
- Kaggle
SafeTune DeepRefusal, LR 2e-5 on Tiny Aya Global
This adapter is a defense that failed the adaptive audit. It is not a hardened deployment model. k*: 1 on both selection seeds, final B150 panel. A0 over-refusal: 110/350 (31.43%).
Model description
Base: Tiny Aya Global, revision af89d219b53ed9b13b8a4645f9c8028973510324, 3.35B parameters, bf16. Method: SafeTune DeepRefusal Variant with a fixed training direction. LoRA rank 16, alpha 16, dropout 0.05, seed 0, learning rate 2e-5, updates 128. Target modules: down_proj, gate_proj, k_proj, o_proj, q_proj, up_proj, v_proj. Training follows SafeTune's defense objective; the audit harness is standalone.
Intended use and limitations
Safety-evaluation research and reproduction of a negative defense result. Transfer-only. Removing the fixed training direction understates the adaptive failure. The measured margin depends on the estimator, seed and panel. It is not an intrinsic safety constant. Judge harm counts are upper bounds, and the article reports capability costs and degeneration separately.
Training data
Training sources include Circuit Breakers and Stanford Alpaca. Alpaca data are CC-BY-NC-4.0; its code license does not relicense the data.
480 rows per class for SGD; 32 reserved training prompts per class for the fixed refusal direction. CuratorKIT dedup and source provenance are recorded in the frozen manifests. Training text and completions are not included in this release. The base license and each upstream data license apply separately.
Load the adapter
Run this on a remote GPU with bf16 support. Use Transformers 5.22.1, PEFT 0.21.2 and Accelerate. First obtain access to the gated base model under Cohere's terms and authenticate with Hugging Face. This example loads the adapter for research; it does not reproduce the audit.
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base_id = "CohereLabs/tiny-aya-global"
base_revision = "af89d219b53ed9b13b8a4645f9c8028973510324"
adapter_id = "Observer04/abliteration-margin-audit-t22"
tokenizer = AutoTokenizer.from_pretrained(base_id, revision=base_revision)
base_model = AutoModelForCausalLM.from_pretrained(
base_id, revision=base_revision, dtype="bfloat16", device_map="auto"
)
model = PeftModel.from_pretrained(base_model, adapter_id, is_trainable=False)
model.eval()
The exact base revision is also stored in adapter_config.json. The adapter repository does not include a tokenizer or the base weights.
Frozen evaluation
| Metric | Value |
|---|---|
| A0 harmful refusal | 123/150 (82.0%) |
| A0 over-refusal | 110/350 (31.43%) |
| Adaptive k* | 1 on both selection seeds, final B150 panel |
| A0 capability | MMLU 269/500; GSM8K 183/250 |
Evidence: aggregate counts and the article. No new evaluation was run for this card.
This repo contains only the defense adapter, its PEFT configuration and this card. The adapter is copied server-side from the frozen training artifact. No abliterated weights, refusal vectors, prompt text or completions are released.
License and attribution
These LoRA weights modify Cohere Labs Tiny Aya Global. Their use and redistribution remain subject to Cohere's CC-BY-NC-4.0 license with Acceptable Use Addendum and Acceptable Use Policy. The adapter is for non-commercial safety research. Keep the base attribution and terms with any redistribution. Upstream training data retain their own terms. The license includes warranty disclaimers. Cohere does not endorse this adapter.
CC-BY-NC-4.0 inherited from CohereLabs Tiny Aya. SafeTune, commit 263ea24d71550dcb8c783226f5f19a7c62f7f7ca. DeepRefusal recipe: Xie et al.. Dual-operator training is a failed adaptation, not a new proven defense.
Citation and contact
@misc{abliteration_margin_audit_2026,
title = {Abliteration Resistance Is a Finite Margin},
year = {2026},
url = {https://github.com/Observer04/abliteration-margin-audit}
}
Report errors or ask reproducibility questions through GitHub issues.
- Downloads last month
- 11