Instructions to use purvansh01/safetypulse-lora-alpaca with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- PEFT
How to use purvansh01/safetypulse-lora-alpaca with PEFT:
Task type is invalid.
- Notebooks
- Google Colab
- Kaggle
SafetyPulse LoRA adapters: Alpaca (safety drift)
Warning. These adapters come from ordinary instruction tuning on Alpaca, and that tuning weakens the model's refusals: attack success rate (ASR) on held-out HarmBench prompts rises from 14.3% (base) to about 25% at the most-drifted checkpoints. They are released as research artifacts for studying and detecting safety drift, not for deployment.
Runs alpaca-s0 (development) and alpaca-s1, alpaca-s2, alpaca-s3 (held-out test seeds), 2,000 Alpaca
examples each. 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.
SafetyPulse's Circuit Integrity Score flagged every seed at step 25, 1–3 checkpoints before ASR rose. SafeTune's safety-vector restore repairs the drift (ASR 25.2% → 12.3%, GSM8K 55.1% → 70.7%).
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-alpaca", subfolder="alpaca-s1/checkpoint-125")
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.
- Downloads last month
- -