Instructions to use RLMIA/GRPO-GLM4-9B-MATH with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use RLMIA/GRPO-GLM4-9B-MATH with Transformers:

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

pipe = pipeline("text-generation", model="RLMIA/GRPO-GLM4-9B-MATH")
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("RLMIA/GRPO-GLM4-9B-MATH")
model = AutoModelForCausalLM.from_pretrained("RLMIA/GRPO-GLM4-9B-MATH", 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 RLMIA/GRPO-GLM4-9B-MATH with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "RLMIA/GRPO-GLM4-9B-MATH"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "RLMIA/GRPO-GLM4-9B-MATH",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/RLMIA/GRPO-GLM4-9B-MATH

SGLang

How to use RLMIA/GRPO-GLM4-9B-MATH 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 "RLMIA/GRPO-GLM4-9B-MATH" \
    --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": "RLMIA/GRPO-GLM4-9B-MATH",
		"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 "RLMIA/GRPO-GLM4-9B-MATH" \
        --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": "RLMIA/GRPO-GLM4-9B-MATH",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use RLMIA/GRPO-GLM4-9B-MATH with Docker Model Runner:
```
docker model run hf.co/RLMIA/GRPO-GLM4-9B-MATH
```

GLM-4-9B MATH GRPO checkpoint

A GLM-4-9B-0414 model fine-tuned on MATH with GRPO for the model-family extension in Auditing Data Membership in Reinforcement Learning With Verifiable Rewards (NDSS 2027).

Reference: zai-org/GLM-4-9B-0414.
Export: the final matched-setting run, iteration 1,139 (1,140 completed training steps; 60 epochs under the recorded schedule).
Format: merged Transformers Glm4ForCausalLM weights, six BF16 safetensors shards and tokenizer.
Recorded training settings: 64 prompts/rollout batch, 8 samples/prompt, global batch size 512, response cap 8,192 tokens, temperature 1.0, learning rate 1e-6, KL coefficient 0.001.
Training input: the 1,269-prompt MATH member split represented by dataset/math-mia/train_half_test_only.parquet in the code repository.

These settings describe this exported checkpoint. They are separate from the audit-time sampling settings.

Code and audit commands: Y-L-LIU/Divergence-in-Behavior-Attack.

This release contains the model and tokenizer only. Optimizer state, intermediate checkpoints, and experiment logs are excluded. Membership-audit performance should be evaluated with the matching member/nonmember data and reference checkpoint.

Downloads last month: 14

Safetensors

Model size

9B params

Tensor type

BF16

Model tree for RLMIA/GRPO-GLM4-9B-MATH

Base model

zai-org/GLM-4-9B-0414

Finetuned

(33)

this model