safetypulse-h2-lora-alpacalist
A set of LoRA checkpoints for studying safety drift
2 min readupdated
A set of 78 LoRA (small add-on to a base model) adapters recorded during the SafetyPulse project experiments based on the CohereLabs/tiny-aya-global language model. The model was fine-tuned on the Alpaca dataset with structured list-style answers, allowing researchers to study safety drift (the weakening of safety refusals during fine-tuning). These checkpoints are intended exclusively for research tasks, vulnerability analysis, and testing safety detectors, rather than for production use.
Key features
- Contains 78 LoRA checkpoints (3 independent runs, steps from 0 to 250)
- Base model: CohereLabs/tiny-aya-global
- Trained on 2,000 Alpaca examples with list-format answers
- PEFT methodology applied (LoRA with r = 16, α = 32)
What you can do
- Use the checkpoints to study safety drift and step-by-step changes in model refusal behavior
- Test vulnerability detection algorithms and monitor fine-tuning processes
- Reproduce SafetyPulse research experiments using open-source code
Example
import torch
from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
BASE, REV = "CohereLabs/tiny-aya-global", "af89d219b53ed9b13b8a4645f9c8028973510324"
tok = AutoTokenizer.from_pretrained(BASE, revision=REV)
base = AutoModelForCausalLM.from_pretrained(BASE, revision=REV, torch_dtype=torch.float16, device_map="auto")
model = PeftModel.from_pretrained(base, "purvansh01/safetypulse-h2-lora-alpacalist",
subfolder="alpacalist-s1/checkpoint-250", revision="h2-v1.1")
Read more
DetailsBenchmarks: peak attack success rate (ASR) increased from 0.21 to 0.46–0.49.
Hardware: training was performed on a Tesla T4 (Kaggle), taking about 51–52 minutes per run.
Limitations: a research artifact, not intended for deployment or production use. CC BY-NC 4.0 license.










Comments