Instructions to use Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT with Transformers:

# Use a pipeline as a high-level helper
from transformers import pipeline

pipe = pipeline("text-generation", model="Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# pip install -U transformers accelerate
# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT")
model = AutoModelForCausalLM.from_pretrained("Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT", device_map="auto")
messages = [
    {"role": "user", "content": "Who are you?"},
]
inputs = tokenizer.apply_chat_template(
	messages,
	add_generation_prompt=True,
	tokenize=True,
	return_dict=True,
	return_tensors="pt",
).to(model.device)

outputs = model.generate(**inputs, max_new_tokens=256)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT

SGLang

How to use Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT with SGLang:

Install from pip and serve model

# Install SGLang from pip:
pip install sglang
# Start the SGLang server:
python3 -m sglang.launch_server \
    --model-path "Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker images

docker run --gpus all \
    --shm-size 32g \
    -p 30000:30000 \
    -v ~/.cache/huggingface:/root/.cache/huggingface \
    --env "HF_TOKEN=<secret>" \
    --ipc=host \
    lmsysorg/sglang:latest \
    python3 -m sglang.launch_server \
        --model-path "Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT with Docker Model Runner:
```
docker model run hf.co/Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT
```

Llama-3.2-1B-Instruct, TOFU forget10, unlearned with GradDiff + LAT

open-unlearning/tofu_Llama-3.2-1B-Instruct_full (Llama-3.2-1B-Instruct fine-tuned on all 4,000 TOFU QA pairs) after unlearning the forget10 split (20 fictitious authors) with GradDiff + LAT.

Note. At these settings this model forgets little of the forget set (see the table): it keeps noticeably more of the forget set than the other unlearned models (see Forget Q-A prob in the table). It is released for completeness.

Algorithm

GradDiff (Gradient Difference; Liu et al., 2022; OpenUnlearning implementation): gradient ascent on the forget set + NLL on the retain set: $\mathcal{L} = -\gamma,\mathrm{NLL}(\text{forget}) + \alpha,\mathrm{NLL}(\text{retain})$.

LAT (Latent Adversarial Training; Sheshadri et al., 2024) is applied on top of GradDiff as a relearning defense: each unlearning step first finds a worst-case per-token $L_2$-bounded perturbation of a hidden layer by PGD that restores the forget answers, then unlearns under that perturbation.

Data

TOFU contains 200 fictitious authors with 20 QA pairs each. forget10 holds 10% of them (20 authors, 400 QA pairs); retain90 holds the other 180 authors. The oracle is the same base model fine-tuned on retain90 only, i.e. a model that never saw the forget set.

Unlearning setup

Hyperparameter	Value
Start model	`open-unlearning/tofu_Llama-3.2-1B-Instruct_full`
Forget set	TOFU `forget10` (400 QA pairs, 20 fictitious authors)
Retain set	TOFU `retain90` (3,600 QA pairs)
Framework	OpenUnlearning (trainer `LATGradDiff`)
Optimizer	paged AdamW (32-bit), weight decay 0.01
Learning rate	1e-05, linear decay, 12 warm-up steps (about 1 epoch)
Batch size	8 x 4 gradient accumulation = 32 (forget), same for retain
Epochs	3
Grad clip	1.0
Precision	fp32 master weights, bf16 autocast
Seed	0
This checkpoint	step 36 (epoch 2.8)
GradDiff gamma (forget weight)	1.0
GradDiff alpha (retain weight)	5.0
GradDiff retain loss	NLL
LAT PGD steps per batch	16
LAT PGD optimizer	AdamW, lr 0.05
LAT perturbation	per-token L2 ball
LAT epsilon	0.5
LAT perturbed layer	model.layers.3 output

Hyperparameter selection. GradDiff and RMU over-forget at the OpenUnlearning defaults and SimNPO under-forgets, so each was tuned with a rule fixed before any attack: over a grid of learning rates (and retain/forget weights), checkpointed every epoch, among checkpoints with mean retain log-probability above -0.5 and forget log-probability at most the retrained model's (-2.58), take the one closest to NPO's forget log-probability (-3.70), near-ties (0.1 nats) broken by retain log-probability. Selected: learning rate 1e-5, retain weight 5, epoch 3. The LAT model reuses the selected GradDiff learning rate, loss weights and length and is taken as trained (final checkpoint), provided its retain log-probability stays above -0.5. The official epsilon of 2 pushes the 1B model's retain log-probability below -0.5; 0.5 was chosen by the same retain rule, before any attack.

Evaluation

OpenUnlearning's TOFU metrics (bf16), computed on the two halves of forget10 (A: even-indexed questions of each author, B: odd-indexed; 200 QA pairs each). Q-A prob (length-normalised probability of the correct answer), ROUGE-L (recall of the greedy answer) and truth ratio are per-question means, so the forget10 value is the mean of the two halves. Forget quality is the p-value of a KS test between this model's truth ratios and the oracle's, reported per half (higher = closer to the oracle). Model utility is the harmonic mean of probability, ROUGE and truth ratio on the retain set, real authors and world facts.

Goal of unlearning: forget metrics move toward the oracle, model utility stays near the starting model.

Model	Forget Q-A prob ↓	Forget ROUGE-L ↓	Forget truth ratio	Forget quality ↑ (half A / half B)	Model utility ↑
TOFU fine-tuned (starting point)	0.880	0.819	0.476	2.0e-15 / 2.4e-10	0.599
Retrained on retain90 (oracle)	0.116	0.378	0.628	1.00 / 1.00	0.592
GradDiff	0.145	0.346	0.513	4.7e-05 / 1.2e-04	0.580
GradDiff-SAM	0.722	0.574	0.494	1.1e-14 / 1.9e-09	0.593
GradDiff-ILU	0.465	0.470	0.513	2.4e-10 / 1.5e-07	0.591
GradDiff-LAT (this model)	0.830	0.682	0.481	4.7e-15 / 2.4e-10	0.592
NPO	0.044	0.275	0.618	0.63 / 0.47	0.576
NPO-SAM	0.026	0.256	0.641	0.11 / 0.63	0.587
NPO-ILU	0.039	0.301	0.644	0.18 / 0.27	0.572
NPO-LAT	0.003	0.156	0.670	0.07 / 0.01	0.582
RMU	0.023	0.185	0.696	0.33 / 0.02	0.567
RMU-SAM	0.204	0.357	0.555	8.1e-08 / 1.8e-05	0.510
RMU-ILU	0.791	0.701	0.511	4.7e-15 / 3.6e-09	0.597
RMU-LAT	0.129	0.335	0.592	4.7e-05 / 3.0e-03	0.529
SimNPO	0.026	0.337	0.622	0.27 / 0.55	0.554
SimNPO-SAM	0.116	0.347	0.630	0.07 / 0.47	0.570
SimNPO-ILU	0.026	0.318	0.619	0.55 / 0.71	0.580
SimNPO-LAT	0.697	0.535	0.517	3.1e-12 / 1.5e-07	0.564

Usage

from transformers import AutoModelForCausalLM, AutoTokenizer
model = AutoModelForCausalLM.from_pretrained("Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT", torch_dtype="auto")
tok = AutoTokenizer.from_pretrained("Lordvarun23/Llama-3.2-1B-Instruct-TOFU-forget10-GradDiff-LAT")