Tiny Aya Global, one-epoch control LoRA adapter

Lexsi AI Safety and Alignment Hackathon, Track 2. Team: Sushovit Nanda and Manish Shaw. Project: Safe from Fine-Tune to Phone. Claim: Benign medical LoRA doubles Tiny Aya Global's attack success (18.2 % to 36.3 %); a run-time steering vector cuts it to 26.3 % and survives GGUF quantization.

All repositories of this project are public and un-gated. No model completion on a harmful prompt is published anywhere; prompt sets carry prompts and labels only, and evaluation logs are text-free records. Write-up, code and reproduction instructions: https://github.com/SushovitNanda/Safe-from-Fine-Tune-to-Phone. Collections: models, datasets.

This repository holds the LoRA adapter trained for one epoch on the token-matched control (Alpaca-cleaned general instructions) set (5,004 rows; dataset SushiOrbit/tinyaya-sft). It is a research artifact built to measure safety drift, not a model for use: Attack success on 856 harmful prompts moves from 18.2 % (original) to 20.0 % with this adapter merged (+1.8 points [-0.9, 4.4]); clinical over-refusal from 2.9 % to 9.5 %. It is the control against which the medical drift is measured.

Recipe: LoRA rank 16, alpha 32, dropout 0.05 on all linear modules (PEFT 0.21.2, Transformers 5.22.1); batch 1 with 16 accumulation steps; AdamW, learning rate 2e-4, cosine schedule, 3 % warm-up, gradient clipping 1.0; FP16 mixed precision; gradient checkpointing; seed 13; loss on assistant tokens only; 30.2 M trainable parameters (0.89 %). Base model: Tiny Aya Global 3.35B at one pinned snapshot (CC-BY-NC-4.0), official chat template with the System Preamble.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", torch_dtype="float16")
model = PeftModel.from_pretrained(base, "SushiOrbit/tinyaya-global-control-lora")
tok = AutoTokenizer.from_pretrained("SushiOrbit/tinyaya-global-control-lora")

Safety note: the merged medical model answers harmful clinical and general requests roughly twice as often as the original in all three languages, most in Hinglish, where most of its answers are also degenerate. Deploy it only with the run-time steering vector of this collection (SushiOrbit/tinyaya-global-refusal-steering-vector) or an equivalent control, and re-measure in the languages of your users. Licence: CC-BY-NC-4.0 (inherited from the base model).

Downloads last month
43
Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for SushiOrbit/tinyaya-global-control-lora

Adapter
(395)
this model

Collection including SushiOrbit/tinyaya-global-control-lora