Instructions to use Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld with libraries, inference providers, notebooks, and local apps. Follow these links to get started.

Libraries

How to use Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld with Transformers:

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

pipe = pipeline("text-generation", model="Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld")
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("Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld")
model = AutoModelForCausalLM.from_pretrained("Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld", 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 Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld with vLLM:

Install from pip and serve model

# Install vLLM from pip:
pip install vllm
# Start the vLLM server:
vllm serve "Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld"
# Call the server using curl (OpenAI-compatible API):
curl -X POST "http://localhost:8000/v1/chat/completions" \
	-H "Content-Type: application/json" \
	--data '{
		"model": "Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Use Docker

docker model run hf.co/Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld

SGLang

How to use Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld 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 "Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld" \
    --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": "Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld",
		"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 "Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld" \
        --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": "Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld",
		"messages": [
			{
				"role": "user",
				"content": "What is the capital of France?"
			}
		]
	}'

Docker Model Runner
How to use Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld with Docker Model Runner:
```
docker model run hf.co/Tim-Xu/Qwen2.5-3B-OPCD-ScienceWorld
```

Qwen2.5-3B-OPCD-ScienceWorld

Qwen2.5-3B-Instruct trained with OPCD on ScienceWorld: on-policy context distillation: the model samples its own response at each training state and the self-teacher scores the sampled tokens.

The training data come from the base model's own ScienceWorld trajectories; no environment is run during training. At evaluation the model acts from the ordinary prompt alone.

Results

ScienceWorld, 384 held-out episodes, in percent:

Policy	Success	Progress
Base policy	6.5	26.3
CD	15.4	46.1
OPCD (this model)	18.5	48.6
CDUP	25.3	52.8

Training

Full fine-tuning of Qwen/Qwen2.5-3B-Instruct on the base model's ScienceWorld deployment states, sampling one response per visit at temperature 1.0.

Evaluation protocol

Seven task types (find-animal, find-living-thing, find-non-living-thing, find-plant, and the three lifespan tasks), easy preset, a five-step history window, a 30-step limit, temperature 0.4.