Instructions to use Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1 with Transformers:

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

pipe = pipeline("text-generation", model="Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1")

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

tokenizer = AutoTokenizer.from_pretrained("Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1")
model = AutoModelForCausalLM.from_pretrained("Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1", device_map="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Use Docker

docker model run hf.co/Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1

SGLang

How to use Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1 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 "Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1" \
    --host 0.0.0.0 \
    --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

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 "Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1" \
        --host 0.0.0.0 \
        --port 30000
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:30000/v1/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1",
		"prompt": "Once upon a time,",
		"max_tokens": 512,
		"temperature": 0.5
	}'

Docker Model Runner
How to use Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1 with Docker Model Runner:
```
docker model run hf.co/Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1
```

Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1

Built with Llama. See LICENSE, NOTICE, and USE_POLICY.md.

An unlearned model from the RAQUEL MUSE experiments: Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-M-orig-LoRA-v1 (M_orig) after LoRA unlearning of the MUSE forget set with SAUL, released at its early-stopped checkpoint. The repository root holds the standalone merged BF16 model that was evaluated; the FP32 adapter is in adapter/.

Method: SAUL: sharpness-aware forget/retain updates with a forget-loss constraint (Lagrange multiplier).
Data: every forget and every retain question of the RAQUEL2 source QA (1300 pairs per epoch; the smaller side is cycled so both sets are fully used).
Schedule: 5 epochs, 205 optimizer steps; a checkpoint was scored every 4 steps.
Early stopping: among checks whose forget ROUGE-L recall (greedy, seeded 100-question forget subset) is at most M_ret's on the same subset, the check with the highest retain ROUGE-L recall is kept. RAQUEL questions were never used for selection. Released checkpoint: step 108 of 205 (forget ROUGE-L 0.068 <= target 0.256; retain ROUGE-L 0.738).

SAUL adaptation. SAUL's sharpness perturbations, its separate forget/retain optimizers and its forget-loss constraint act on the LoRA adapter weights rather than on the full model (parameter_space: lora_adapters); treat it as a LoRA adaptation of the published algorithm.

Evaluation

Split	This model	M_orig	M_ret
Forget (original)	17/750 (2.3%)	99.7%	23.5%
Forget (paraphrased)	22/741 (3.0%)	91.1%	21.5%
Retain (original)	814/1300 (62.6%)	100.0%	100.0%
Retain (paraphrased)	649/1283 (50.6%)	87.5%	88.9%
RAQUEL affected	14/1709 (0.8%)	11.8%	9.0%
RAQUEL unaffected	51/2010 (2.5%)	16.0%	12.7%

Semantic accuracy judged by Qwen/Qwen3.8-27B (vLLM 0.28.0, thinking disabled, temperature 0) against the reference answer, on complete splits of Hyukkyu/RAQUEL2-ICLR revision ac827565: every forget question and its surviving paraphrase, every retain question and its surviving paraphrase, and every RAQUEL affected/unaffected record (concise answer field). Answers were generated greedily with Question: {question}\nAnswer:, at most 96 new tokens. Per-split counts and evidence hashes are in evaluation.json. M_orig and M_ret rows are the reference baselines on the same splits.

Training

Start: Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-M-orig-LoRA-v1, revision 01a00ac5c754503250e4aa0e1492ab043e52634a.
LoRA rank 64, alpha 128, dropout 0.05 on q/k/v/o/gate/up/down projections; BF16 base, FP32 adapters; one GPU.
Learning rate None (10x the full-parameter protocol), constant schedule; global batch 32; seed 0; max length 512.
Method settings: see training_recipe.json.
Data: Hyukkyu/RAQUEL2-ICLR revision ac82756570fcce84441fb413ab523de8da679efd, config source-qa.

Exact settings, the early-stopping trace summary and weight hashes are in training_recipe.json.

Load

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

repo_id = "Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1"
tokenizer = AutoTokenizer.from_pretrained(repo_id)
model = AutoModelForCausalLM.from_pretrained(repo_id, dtype=torch.bfloat16, device_map="auto").eval()
prompt = "Question: {question}\nAnswer:"

The root merged weights are the evaluated artifact. The FP32 LoRA adapter is in adapter/ (PeftModel.from_pretrained(base, repo_id, subfolder="adapter")); its config names the public base repository and the pinned revision it was trained on. Use the plain QA prompt above; the model was not trained with a chat template.

Downloads last month: 243

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1

Base model

meta-llama/Llama-3.1-8B

Finetuned

Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-M-orig-LoRA-v1

Finetuned

(5)

this model

Hyukkyu
/

Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1

Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1

Evaluation

Training

Load

Model tree for Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1

Datasets used to train Hyukkyu/Llama-3.1-8B-RAQUEL-MUSE-Unlearn-SAUL-LoRA-v1