hubble-1b-npo-unlearned

Base model : allegrolab/hubble-1b-100b_toks-perturbed-hf @ step48000
Method : NPO (Negative Preference Optimization) + LoRA
Forget set : allegrolab/biographies_yago dup >= 64 (268 examples)
Retain set : dup < 64 (2232 examples)

LoRA config

• r=16, lora_alpha=32, dropout=0.05
• target_modules: q_proj, v_proj
• trainable params: ~1.7M / 1.18B (0.14%)

Hyperparameters

lr=5e-05, β=0.1, α=1.0, effective_batch=8

Training notes

Stopped at epoch 3 step 190 / 5 epochs. Forget loss: 13.86 → ~9.5 (meaningful unlearning signal). Retain loss: stable ~1.7 epochs 1-2, rising ~3.0 at stop point.

Loading

from transformers import AutoModelForCausalLM
from peft import PeftModel

base = AutoModelForCausalLM.from_pretrained(
    "allegrolab/hubble-1b-100b_toks-perturbed-hf", revision="step48000", torch_dtype=torch.float16
)
model_npo = PeftModel.from_pretrained(base, "Harsh01012/hubble-1b-npo-unlearned")