Instructions to use ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final with Transformers:

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

pipe = pipeline("text-generation", model="ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final")
messages = [
    {"role": "user", "content": "Who are you?"},
]
pipe(messages)

# Load model directly
from transformers import AutoTokenizer, AutoModelForCausalLM

tokenizer = AutoTokenizer.from_pretrained("ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final")
model = AutoModelForCausalLM.from_pretrained("ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final", 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 ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final

SGLang

How to use ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final 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 "ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final" \
    --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": "ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final",
		"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 "ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final" \
        --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": "ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final with Docker Model Runner:
```
docker model run hf.co/ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final
```

Phase 2 Final Model

Merged full model produced by the Phase 2 online-DPO / SDPO-style experiment in the Lab 0 self-training project.

Fair evaluation

Deterministic greedy exact-match evaluation with the required #### <number> answer parser:

Full-set eval: 45.41% GSM8K / 57.67% SVAMP
Second fixed-subset eval: 48% GSM8K / 50% SVAMP

These results currently outperform the old Phase 3 REINFORCE checkpoints under the same fair-eval setup.

Notes

Base model: Qwen/Qwen2.5-1.5B-Instruct
Hardware/training constraints: fp16 + SDPA, no bf16, no Flash Attention 2
This is a research artifact, not a production math model

Extra files

phase2_final_full_eval.json
phase2_vs_phase3_independent_shuffle.json

Downloads last month: 6

Safetensors

Model size

2B params

Tensor type

F32

Model tree for ahmed-3m/qwen25-1.5b-gsm8k-sdpo-final

Base model

Qwen/Qwen2.5-1.5B

Finetuned

Qwen/Qwen2.5-1.5B-Instruct

Finetuned

(1800)

this model