Instructions to use PeterWright/Qwen2.5-1.5B-MathHard-GRPO with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use PeterWright/Qwen2.5-1.5B-MathHard-GRPO with PEFT:

from peft import PeftModel
from transformers import AutoModelForCausalLM

base_model = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Math-1.5B-Instruct")
model = PeftModel.from_pretrained(base_model, "PeterWright/Qwen2.5-1.5B-MathHard-GRPO")

Transformers

How to use PeterWright/Qwen2.5-1.5B-MathHard-GRPO with Transformers:

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

pipe = pipeline("text-generation", model="PeterWright/Qwen2.5-1.5B-MathHard-GRPO")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoModel
model = AutoModel.from_pretrained("PeterWright/Qwen2.5-1.5B-MathHard-GRPO", dtype="auto")

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use PeterWright/Qwen2.5-1.5B-MathHard-GRPO with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "PeterWright/Qwen2.5-1.5B-MathHard-GRPO"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "PeterWright/Qwen2.5-1.5B-MathHard-GRPO",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/PeterWright/Qwen2.5-1.5B-MathHard-GRPO

SGLang

How to use PeterWright/Qwen2.5-1.5B-MathHard-GRPO 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 "PeterWright/Qwen2.5-1.5B-MathHard-GRPO" \
    --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": "PeterWright/Qwen2.5-1.5B-MathHard-GRPO",
		"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 "PeterWright/Qwen2.5-1.5B-MathHard-GRPO" \
        --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": "PeterWright/Qwen2.5-1.5B-MathHard-GRPO",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use PeterWright/Qwen2.5-1.5B-MathHard-GRPO with Docker Model Runner:
```
docker model run hf.co/PeterWright/Qwen2.5-1.5B-MathHard-GRPO
```

Qwen2.5-Math-1.5B-Instruct — MathHard GRPO LoRA Adapter

LoRA adapter trained on the Hendrycks MATH level-5 subset using GRPO (Group Relative Policy Optimization).

Training Details

Base model: Qwen/Qwen2.5-Math-1.5B-Instruct
Algorithm: GRPO (PPO-style clipped surrogate, group-relative advantage normalization)
Dataset: the-jb/hendrycks-math — level 5 (hardest) problems only
Steps: 501
Batch size: 4 prompts × 4 samples (group size) = 16 rollouts per step
PPO epochs: 2
Max new tokens: 512
Learning rate: 3e-5
KL coefficient: 0.05
Clip epsilon: 0.2

LoRA Config

Parameter	Value
Rank	16
Alpha	32
Dropout	0.05
Target modules	q/k/v/o_proj, gate/up/down_proj
Trainable params	~18.5M / 1562M total

Reward Function

Correct reward: 1.0 — exact numeric match inside \boxed{}
Format reward: 0.1 — for including \boxed{} in output

Evaluation Results (held-out test set, 512 level-5 problems)

Metric	Baseline (step 0)	Final (step 501)
Exact match (boxed parser)	22.7%	35.7%
Exact match (relaxed parser)	26.2%	38.7%

Usage

from peft import PeftModel
from transformers import AutoModelForCausalLM, AutoTokenizer

base = AutoModelForCausalLM.from_pretrained("Qwen/Qwen2.5-Math-1.5B-Instruct")
model = PeftModel.from_pretrained(base, "PeterWright/Qwen2.5-1.5B-MathHard-GRPO")
tokenizer = AutoTokenizer.from_pretrained("PeterWright/Qwen2.5-1.5B-MathHard-GRPO")

messages = [
    {"role": "system", "content": "Solve the competition-level math problem.\nReturn the final answer in this exact format: \\boxed{NUMBER}"},
    {"role": "user", "content": "What is the sum of all integers from 1 to 100?\n\nReturn only: \\boxed{NUMBER}"}
]
text = tokenizer.apply_chat_template(messages, tokenize=False, add_generation_prompt=True)
inputs = tokenizer(text, return_tensors="pt")
outputs = model.generate(**inputs, max_new_tokens=512)
print(tokenizer.decode(outputs[0], skip_special_tokens=True))

Framework versions

PEFT 0.19.1

Downloads last month: 14

Model tree for PeterWright/Qwen2.5-1.5B-MathHard-GRPO

Base model

Qwen/Qwen2.5-1.5B

Finetuned

Qwen/Qwen2.5-Math-1.5B

Finetuned

Qwen/Qwen2.5-Math-1.5B-Instruct

Adapter

(15)

this model