Instructions to use Jeesup/svd-safety-l2_swift_jbbresp13_remove50 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Jeesup/svd-safety-l2_swift_jbbresp13_remove50 with Transformers:

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

pipe = pipeline("text-generation", model="Jeesup/svd-safety-l2_swift_jbbresp13_remove50")
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("Jeesup/svd-safety-l2_swift_jbbresp13_remove50")
model = AutoModelForCausalLM.from_pretrained("Jeesup/svd-safety-l2_swift_jbbresp13_remove50", 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 Jeesup/svd-safety-l2_swift_jbbresp13_remove50 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Jeesup/svd-safety-l2_swift_jbbresp13_remove50"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Jeesup/svd-safety-l2_swift_jbbresp13_remove50",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Jeesup/svd-safety-l2_swift_jbbresp13_remove50

SGLang

How to use Jeesup/svd-safety-l2_swift_jbbresp13_remove50 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 "Jeesup/svd-safety-l2_swift_jbbresp13_remove50" \
    --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": "Jeesup/svd-safety-l2_swift_jbbresp13_remove50",
		"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 "Jeesup/svd-safety-l2_swift_jbbresp13_remove50" \
        --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": "Jeesup/svd-safety-l2_swift_jbbresp13_remove50",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Jeesup/svd-safety-l2_swift_jbbresp13_remove50 with Docker Model Runner:
```
docker model run hf.co/Jeesup/svd-safety-l2_swift_jbbresp13_remove50
```

svd-safety-l2_swift_jbbresp13_remove50

Built with Llama 2.

Swift-SVD compression of meta-llama/Llama-2-7b-chat-hf, targeting 50% parameter removal. Calibration: 256 WikiText sequences + 13 JBB harmful-prompt/refusal sequences, 2048 tokens each, seed 42. The extra corpus is jbb_harm_resp:13 (26,624 tokens). Both model families use the existing Llama-2 refusal texts, packed with each family's own tokenizer. This is prompt-plus-response calibration, not prompt-only.

Recovery: SVD-LLM recovery: sequential LoRA on U then V factors, plain Alpaca-cleaned data, two epochs per training pass. Dynamic rank allocation uses alpha 0.6; layer importance remains WikiText-based. The checkpoint is folded to standard dense Transformers weights: the compression ratio describes the factorized parameter budget, not a reduction in download size. Realized parameter counts and recovery details are recorded in compression.json.

Evaluation

Full WikiText-2 perplexity; seven-task zero-shot ARC-Easy/Challenge, HellaSwag, WinoGrande, OpenBookQA, PIQA and MathQA; XSTest-safe and OR-Bench over-refusal judged by WildGuard; full AdvBench and StrongREJECT prompt sets judged by HarmBench. StrongREJECT uses the HarmBench judge, not its native scorer. Generation limit: 512 tokens; no smoke-test sample caps.

Metric	Value
WikiText-2 perplexity	14.2221
AdvBench ASR (HarmBench)	0.0269231
StrongREJECT ASR (HarmBench)	0.0351438
Macro over-refusal (WildGuard)	0.536183

All raw evaluation files, including per-prompt generations, judge records, zero-shot sample logs and summaries, are included in utility/ and safety/. Read these files for coverage and judge reliability. This research checkpoint is intended for measuring safety/utility under compression; these measurements do not establish suitability for deployment.

Downloads last month: 277

Safetensors

Model size

7B params

Tensor type

BF16

Model tree for Jeesup/svd-safety-l2_swift_jbbresp13_remove50

Base model

meta-llama/Llama-2-7b-chat-hf

Finetuned

(749)

this model