Instructions to use WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2 with Transformers:

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

pipe = pipeline("text-generation", model="WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2")
model = AutoModelForCausalLM.from_pretrained("WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2", 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=40)
print(tokenizer.decode(outputs[0][inputs["input_ids"].shape[-1]:]))

Notebooks
Google Colab
Kaggle
Local Apps Settings

vLLM

How to use WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2 with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2

SGLang

How to use WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2 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 "WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2" \
    --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": "WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2",
		"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 "WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2" \
        --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": "WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2 with Docker Model Runner:
```
docker model run hf.co/WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2
```

Qwen2.5-7B-Instruct · student-likeness conditional error generator (GRPO, epoch 2.0)

Given a math problem and a misconception description, the model writes a student-style solution that makes that error and ends with a (wrong) final answer. Trained with GRPO (TRL 1.14) from Qwen/Qwen2.5-7B-Instruct. Run student_likeness_seed42_20260929_115310, snapshot epoch-2.0 (step 682 of 682, the final model); it had the highest mean test reward of the five evaluated models.

Input format (use exactly this)

Chat template of Qwen2.5-Instruct with two messages:

system: the text below with {error_description} replaced by the misconception description
user: the problem text

You are a student who makes the following mathematical error.
Solve the problem in a way that naturally reflects this error.

Error description:
{error_description}

Show your working and give your final answer.
Write the answer itself, not an option letter or option number.
Write as that student would, believing your working is correct.
Do not mention the error, do not describe what you are doing wrong,
and do not state or compare with the correct answer.

Sampling used in training and evaluation: temperature 1.0, top_p 1.0, no top-k (0), repetition penalty 1.0, max 1024 new tokens, stop on <|im_end|> / <|endoftext|>. generation_config.json holds exactly these settings, so generate() and vLLM defaults sample like the RL run.

Training

Data: Eedi misconception pairs (KEEP problems), split 80:20 by problem text; train 2,048 pairs / 965 questions.
Reward per rollout = main + 0.5 × student-likeness + truncation
- main: +1 if the final answer is wrong (gpt-5-nano extraction and grading) and the reward verifier (WooYoungSeok/qwen2.5-math-7b-descriptive-verifier-v2-trval-halfA) labels both of 2 samples (T 0.6) aligned with the misconception; 0 if wrong but not aligned; −0.75 if correct or no final answer
- student-likeness: normalized pairwise win rate among the accepted rollouts of a group (gpt-5-nano judge, 2 real MathEDU student solutions as style references)
- truncation: −0.5 when cut off without EOS
GRPO: 8 rollouts per prompt, 48 rollouts (6 prompts) per optimizer step, 682 steps (2 epochs), lr 1e-6 with 10% warmup and linear decay, beta 0.04, epsilon 0.2, loss dapo, rewards scaled per group.

Test results (481 held-out pairs × 8 rollouts)

Scored like the training reward, but with the held-out verifier WooYoungSeok/deepseek-r1-0528-qwen3-8b-descriptive-verifier-v2-trval-halfB. The snapshot was chosen on this test set, so its scores are optimistic.

metric	base Qwen2.5-7B-Instruct	this model
mean reward	0.088	0.760
success rate (wrong answer and both verifier samples aligned)	41.8%	75.0%
correct-answer rate	57.4%	23.6%
wrong answers accepted by the verifier	98.1%	98.2%
wrong answers equal to the target distractor	24.7%	21.2%

Caveat: the test verifier accepts almost every wrong answer (base model included), and it separates misconceptions of the same question poorly. So the success rate mostly tracks how often the model answers wrongly, and it is weak evidence that the specific target misconception is followed.

Downloads last month: 5

Safetensors

Model size

8B params

Tensor type

BF16

Model tree for WooYoungSeok/qwen2.5-7b-instruct-student-likeness-error-generator-epoch2

Base model

Qwen/Qwen2.5-7B

Finetuned

Qwen/Qwen2.5-7B-Instruct

Finetuned

(3142)

this model