SafetyPulse LoRA adapters: negative control

Runs none-s0 (development), none-s1, none-s2 (test): 1,800 Alpaca examples plus 200 refusals to harmful requests (safety-preserving data). LoRA r = 16 on all attention and MLP projections, lr 2e-4, 2 epochs, effective batch 16, about 250 steps, a checkpoint every 25 steps (11 per run, checkpoint-0 is the untouched base). Each folder <run>/checkpoint-<step>/ holds adapter_config.json and adapter_model.safetensors.

Negative control: ASR fell (test 6.7%) and the monitor stayed PASS at every checkpoint.

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
base = AutoModelForCausalLM.from_pretrained("CohereLabs/tiny-aya-global", torch_dtype="auto")
model = PeftModel.from_pretrained(base, "purvansh01/safetypulse-lora-none", subfolder="none-s1/checkpoint-250")

Part of SafetyPulse (Lexsi Labs AI Safety & Alignment Hackathon, Track 2b): monitoring safety drift during fine-tuning through the model's refusal mechanism. Licence CC-BY-NC-4.0, inherited from Tiny Aya and Alpaca. Base model: CohereLabs/tiny-aya-global.