Tiny Aya Water Multilingual Safety LoRA
adapters for researching multilingual LLM safety
2 min readupdated
A collection of LoRA adapters (lightweight add-ons for a base model) and behavior-steering vectors built to study language model safety across different languages. Based on CohereLabs/tiny-aya-water, the project accompanies a research paper investigating whether safety fine-tuning actually makes a model safer or simply triggers more refusals. The repository includes adapters trained on the exact same 568 examples in English, Japanese, and both languages combined, alongside control variants with different training lengths. These resources are aimed at AI safety researchers exploring cross-lingual constraint transfer and the over-refusal problem.
Key features
- Training targets only a single layer,
model.layers.18.mlp.down_proj(just 208,896 trainable parameters at rank 16 and alpha 32). - Multiple training configurations organized in subfolders: monolingual (
en,ja), bilingual (bi,bi_h2), and control runs across three different seeds. - Weight-space steering directions (in
.npzformat) included in thesteering_directions/folder to mitigate over-refusals and harmful outputs. - Open reproducibility: the author released the training data streams, model outputs, and automated judge evaluations.
What you can do
- Test in practice how safety fine-tuning in one language affects refusal rates in another (the cross-lingual effect).
- Apply the precomputed steering directions to the base model's
o_projanddown_projmatrices to reduce false refusals on benign prompts without full retraining.
Example
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-water", torch_dtype="bfloat16")
tok = AutoTokenizer.from_pretrained("CohereLabs/tiny-aya-water")
model = PeftModel.from_pretrained(base, "isaac369369/tiny-aya-water-multilingual-safety-lora", subfolder="ja/seed0")Read more
DetailsMetrics (evaluated via Qwen3Guard 0.6B): the base model produces 11.13% unsafe responses in English and 15.37% in Japanese, with false refusal rates of 4.26% and 3.32%.
The bilingual
bi adapter drops unsafe responses to 4.03% (EN) and 9.68% (JA), but pushes false refusals up to 36.00% and 15.35%.Training setup: PEFT LoRA (rank 16, alpha 32, dropout 0.05), AdamW 1e-4, 1 epoch, batch size 1, max length 768.
Limitations: labels were generated automatically by Qwen3Guard without human verification; licensed under CC BY-NC 4.0 (research only).










Comments